AI SearchAnswers.

An engine that is about to name three vendors has to be confident the thing it is naming exists, does what the user needs, and can be described without contradiction. It builds that confidence from your site, your structured data and independent sources that discuss you.

When those three disagree, or when the last of them is simply absent, the safest move for the model is to name a competitor it can verify. Most invisibility cases we open are not a content problem at all. They are an identity problem: the same company described four different ways across its own pages, with no machine-readable link between them.

The overlap is larger than most vendors admit. Google's own guidance on its generative experiences, updated 10 July 2026, still lists conventional SEO, genuinely useful content and technical accessibility as the foundation. There is no separate mandatory GEO technology and no structural shortcut.

What actually changes is the unit of success. In SEO you win a position. In GEO you win a sentence, and sometimes a link next to it. That shifts the work toward clear claims, clean entity data and sources an engine can stand behind.

The longer comparison is here.

It should never be prioritized over pages that genuinely answer a question. If an agency is leading its pitch with llms.txt as a headline deliverable, that is a useful signal about the rest of the proposal.

Cipion publishes one at cipionmarketing.com/llms.txt on the same logic: cheap, harmless, not the point.

In practice that means three things. Answer the question in the first 40 to 60 words of the page, before any preamble, because that block is what gets lifted. Use the wording real people use, not internal product language. And make the claim checkable, with a date, a number or a named source, because an engine summarizing for a stranger prefers a sentence it can stand behind.

What does not work: keyword stuffing, publishing the same page for forty cities, or adding schema to a page that says nothing. Google's spam policies explicitly cover scaled content abuse, and a thin page with perfect markup is still a thin page.

Worth setting expectations honestly: appearing in an AI Overview is not something anyone can guarantee, and the set of citations for a given query changes over time. The longer guide is here.

Neither acronym is used by the platforms themselves. Google's guidance for its generative experiences, updated 10 July 2026, still describes conventional SEO, genuinely useful content and technical accessibility as the foundation. There is no AEO specification and no mandatory GEO technology. Anyone selling you one of the two labels as a distinct product owes you an explanation of what is actually different in the work.

Where a real difference shows up is in measurement. The older AEO playbook aimed at a snippet you could open, see and screenshot. A generated answer is assembled per query and varies between runs, so the honest unit is frequency of mention across repeated runs of a fixed set of prompts, not a position. On the Google side there is now one native signal: the generative AI performance report in Search Console, whose worldwide rollout was announced on 31 August 2026, gives impressions from AI Overviews and AI Mode by page, country, date and device. It does not give clicks or conversions.

So the useful question is not which acronym an agency prefers. It is what they will measure, how often they will rerun it, and what the baseline looked like before the work started. The longer AEO walkthrough is here.

For Google the bar is documented and short. Its AI features guide says a page must be indexed and eligible to be shown in Google Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional technical requirements. So noindex, nosnippet, data-nosnippet and a tight max-snippet all limit what can be shown. Search Console also has a Search generative AI control. It includes your site by default, but a property inherits the setting from its closest parent that stopped inheriting, so check it on every property you own before blaming the content.

ChatGPT and Perplexity each run their own search crawler. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, although they can still appear as navigational links. GPTBot is a separate agent used for model training, so blocking it does not remove you from ChatGPT search. Perplexity says PerplexityBot surfaces and links websites in its results, and recommends allowing it in robots.txt and permitting its published IP ranges. A security plugin, CDN bot rule or hosting firewall can block these crawlers even when robots.txt allows them, so check your server and CDN logs for refused requests from those agents.

If access is clean and you are still absent, the problem has moved up a layer: the engines can reach you but have no reason to name you. That is an entity and content question, covered in the first answer on this page, and it takes longer. Fixing access makes you eligible, not chosen, and nobody can honestly promise when or whether an engine will cite you. When we audited our own domain in September 2026, this layer was where the first findings were, including a robots.txt that allowed GPTBot but not OAI-SearchBot. The full list is in how to program an agent to run your GEO maintenance.

Google's own documentation for AI Overviews and AI Mode, last updated on 10 December 2025, states that a page only needs to be indexed and eligible to show with a snippet, with no additional technical requirements, and it lists keeping Business Profile information up to date among the basics. Google's local ranking help page names the three factors behind local results: relevance, distance and prominence, and says plainly that there is no way to request or pay for a better local ranking. A local business controls relevance through accurate categories, services and page content, and influences prominence through the reviews, links and coverage it earns. Distance depends on where the searcher is, so no provider can guarantee that a local business appears for every nearby query, in the map results or in an AI answer.

Google Business Profile guidelines set the boundary on what a local business can claim. A profile is only allowed for a business that has a physical location customers can visit or that travels to its customers, and a service-area business cannot list a virtual office unless that office is staffed during business hours. Creating profiles or near-identical city pages in towns where the business has no real presence breaks those guidelines, and mass-produced location pages can also fall under Google's spam policy on scaled content abuse. The honest local playbook is narrower: one accurate profile per real location, one website page for each service the business actually offers, and the same name, address and phone number wherever the business is listed.

The main difference between local and national GEO is the unit of measurement. A national brand tracks whether it is named across broad category prompts, while a local business should track a short list of prompts that include its city or neighborhood, such as a service plus a place name, and log the answers every month with the date and the platform. Answers to local prompts also shift with the searcher's location, so a single test from one address is not a baseline. Search Console's generative AI report shows impressions by page, country, date and device, not by city or by query, so for a local business the prompt log remains the main evidence. A sensible first step is a technical and entity review like the one in our GEO audit checklist.

Google's documentation, last updated on 10 December 2025 and read on 30 September 2026, says a page must be indexed and eligible to be shown in Search with a snippet, and that there are no additional requirements or special optimizations for AI Overviews or AI Mode. Google lists three practical basics: allow crawling in robots.txt, make content findable through internal links, and keep important content in textual form. Google also names the controls that limit appearance: robots.txt directives for Googlebot, preview controls such as nosnippet and max-snippet, and noindex. For Google, the work is ordinary search hygiene, and the AI features sit on top of it.

OpenAI's crawler documentation says OAI-SearchBot surfaces sites in ChatGPT search, and a site that blocks it will not be shown in ChatGPT search answers, though it can still appear as a navigational link. OpenAI says robots.txt changes take about 24 hours to apply and recommends allowing its published IP ranges. GPTBot is a separate crawler used for model training, so allowing search does not require allowing training. Perplexity's documentation says PerplexityBot exists to surface and link sites in Perplexity results, is not used to train foundation models, and should be allowed in robots.txt along with its published IP ranges. Perplexity-User visits pages when a person asks a question.

The shared groundwork across all three platforms is a site that can be crawled, pages that are indexed where an index applies, and copy that states plainly who the business is and what it offers. No platform documents a mechanism that turns eligibility into a citation, so measurement should be a dated log of the same prompts on each platform, reported as observations and never as guaranteed results. A technical and entity review like the one in our GEO audit checklist is a sensible first step, and the answer on brand invisibility above covers what to check when access is fine and the brand still does not show up.

A dated baseline comes first, because a fix without a before measurement cannot be judged. Keep a log of the prompts a buyer would actually type, run each one in a fresh session on ChatGPT, Perplexity and Google, and record the date, whether your brand appeared and whether the description was accurate. Search Console's generative AI report adds impressions from AI Overviews and AI Mode by page, country, date and device, without clicks or queries. With the baseline saved, check access: Google's AI features documentation, last updated 10 December 2025, says a page only needs to be indexed and eligible for a snippet, with no additional technical requirements. Confirm that robots.txt allows OAI-SearchBot and PerplexityBot and that no firewall rule blocks them. The answer above on brands that do not appear at all covers each of those checks in detail.

Domain cleanup is the step most teams skip, and it often matters more than new schema. Crawl every URL your domain actually serves, not just the pages in your navigation: forgotten CMS installs, demo posts, tag archives, feeds and old landing pages all describe your brand to a crawler. On our own domain in September 2026, an audit found a hidden WordPress install serving 28 theme demo posts under the Cipion name, with archives, feeds and a REST API repeating them. Retiring that material with 410 responses came before any new content, as described in our guide to running GEO maintenance with an AI agent. Google's spam policies, updated 28 August 2026, define scaled content abuse as many pages generated mainly to manipulate rankings rather than to help users, so the fix is fewer, truer pages.

Entity alignment comes next. Your name, description, services and location should read the same on your site, in your structured data and on the third party profiles you control, and Google lists structured data that matches the visible text as a best practice for its AI features. Link your Organization and Person entities by @id and point sameAs only at profiles that really belong to you. New content comes last, written as direct answers to questions buyers already type, which Search Console queries reveal. Then rerun the same prompt log with identical wording and compare the two dates. No platform documents a timeline or a way to guarantee a citation, so report what changed as dated observations with their sources.

OpenAI's crawler documentation, read on 6 October 2026, separates three agents. OAI-SearchBot is the one used to surface websites in ChatGPT's search features, and OpenAI says sites that opt out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links. GPTBot is a different agent used for training foundation models, so blocking GPTBot does not remove you from ChatGPT search. OpenAI also notes that a robots.txt change takes about 24 hours to reach its systems. The first practical check is therefore your robots.txt and your server logs: OAI-SearchBot has to be allowed, and it has to receive a normal 200 response, because a rule that allows the bot is useless if the server rate limits it.

Once access is clean, the work moves to what ChatGPT can say about you. ChatGPT search answers questions by retrieving pages and summarising them, so the pages that get used are the ones that answer a specific question in plain language near the top, with the facts and their source in the same paragraph. Vague service pages that never name the problem, the customer or the location give the system little to quote. A useful exercise is to write down the ten questions a buyer would ask ChatGPT before hiring you, run them, and log which sources appear. Those sources show you the format ChatGPT is already choosing for your category.

Ranking language can mislead here. A ChatGPT answer is generated per question and can change between sessions, so there is no stable position one to hold. Measurement means repeating the same prompts on a fixed schedule and logging the date, the answer and the cited links, then judging the trend over weeks. Anyone who promises a guaranteed spot in ChatGPT is promising something OpenAI does not offer. The groundwork we check first is listed in our OAI-SearchBot vs GPTBot guide.

Start with access, because it is the cheapest cause to rule out. OpenAI's documentation, read on 6 October 2026, says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, and that GPTBot, the training crawler, is a separate agent. A site that blocked GPTBot may have copied the same rule to every OpenAI agent, and some firewalls and hosting plans challenge bots they do not recognise. If Googlebot is allowed while OAI-SearchBot is challenged, Google rankings stay healthy while ChatGPT never sees the pages. A request that sends the OAI-SearchBot user agent to your home page should return a normal 200, and your server logs should show the bot visiting.

If access is clean, the second cause is format. Google can rank a page for its overall relevance and authority, while ChatGPT search tends to lift a specific passage that answers the question asked. A service page that ranks on brand strength but never states plainly what the business does, for whom and where, gives ChatGPT little it can quote. Rewriting the opening of each key page so the first two sentences answer the buyer's question is often the single most useful content change.

The third cause is outside your site. When ChatGPT answers a recommendation question, it draws on many sources, and a business that appears only on its own website is easy to leave out. Consistent descriptions in directories, reviews, industry lists and press help the engine connect your name with your category. None of these steps guarantees a mention, but together they remove the reasons for being skipped. The checks behind each of these causes are part of our GEO audit.

The crawler documentation of each company, read on 6 October 2026, makes the split explicit. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers. Perplexity says PerplexityBot is designed to surface and link websites in its search results and is not used for training foundation models. Anthropic says blocking Claude-SearchBot prevents indexing for search and may reduce visibility in its results. Google says a page must be indexed and eligible for a snippet to appear in AI Overviews. A robots.txt rule, firewall or hosting filter can allow one of these agents and block another, which produces exactly the pattern of present in one engine and absent in the next.

When access is identical, the difference comes from sources. Each engine retrieves from its own index and weighs pages in its own way, so the pages it trusts for your category can differ. One engine may favour directories and review sites, another may lean on recent articles, another on the brand's own pages. Running the same ten buyer questions in every engine and logging the cited URLs shows which sources each one prefers, and those sources are where your business needs to be described accurately and consistently.

Timing explains some gaps too. Crawlers revisit sites at different rates, OpenAI notes that robots.txt changes take about 24 hours to reach its systems, and answers can change between sessions. A brand that appears in one engine today and disappears tomorrow is more likely seeing normal variation than a penalty. Judge the pattern over several dated rounds before changing anything, and check the server logs for each crawler first, since a missing visit is the simplest explanation and the quickest one to fix. Our robots.txt guide for AI crawlers compares the agents of each company in one table.

Perplexity's crawler documentation, read on 7 October 2026, describes two agents. PerplexityBot is the one that surfaces and links websites in Perplexity search results, and Perplexity states it is not used to crawl content for AI foundation models. Perplexity recommends allowing PerplexityBot in robots.txt and permitting requests from its published IP ranges, and it gives step by step rules for Cloudflare and AWS firewalls, because a security layer can block the bot even when robots.txt allows it. The second agent, Perplexity-User, fetches a page when a person asks a question, and Perplexity says it generally ignores robots.txt because a user requested the fetch. Changes to these settings can take up to 24 hours to reach Perplexity's systems, so test access a day after any edit.

Perplexity's help center article on how the product works, last modified on 3 September 2026, says it searches the web in real time, summarises what it finds, and attaches numbered citations that link to the original sources. Its Research mode runs dozens of searches and reads hundreds of sources for a single report. For a business, that means each citation is earned page by page and question by question. The pages that get quoted tend to state the answer in the first lines, name the product, the customer and the location in full, and keep each fact next to its source. A service page that never names the problem it solves gives Perplexity very little to lift into an answer.

Ranking is the wrong frame for Perplexity because the answer is generated for each question and the cited sources can change between runs, between models and between Pro Search and a standard search. A sensible routine is to pick the ten questions a buyer would ask before hiring you, run them in Perplexity on a fixed schedule, and log the date, the answer and every cited link. Over several weeks that log shows which formats and which domains Perplexity prefers in your category. Perplexity does not sell a place among its cited sources, so treat any promise of a fixed spot with suspicion. The access checks we run first, across Perplexity, OpenAI and Google crawlers, are in our AI crawler robots.txt guide.

It reports impressions from AI Overviews and AI Mode, broken down by page, country, date and device. That is a real measurement and a genuine step forward.

What it does not give you: clicks or conversions attributable to AI answers. The data is also counted inside your Web totals rather than as a separate audience, so it cannot be added on top of existing numbers. Before promising a client this report, confirm the property actually has it and has enough impressions to be worth reading.

Citation share is the percentage of your tracked prompts where your domain is cited at all. Answer presence is whether the brand is named even without a link, which matters because ChatGPT often does exactly that. Position is where you sit in the citation list when you are there.

Anything presented to you as an exact "AI traffic" or "AI revenue" figure is an estimate built on assumptions. Estimates are fine. Estimates presented as measurements are not.

Build a fixed set of 20 to 40 prompts covering your category, run them on a schedule, and record which domains are cited, in what position and with what framing. Keep the set fixed, because changing the prompts between runs destroys the comparison.

That log is the only honest baseline for AI visibility work. It measures what the engine actually returned, on a date, rather than what a tool estimated.

The practical consequence is asymmetric. Perplexity is easier to audit, because the sources are right there and a click can produce measurable referral traffic. ChatGPT is easier to underestimate, because a brand can shape the answer a buyer reads and generate almost no trackable click while doing it.

If you judge AI search purely by referral traffic in analytics, you will conclude ChatGPT does nothing for you. That conclusion is usually wrong.

What makes the log longitudinal is what stays constant. Same prompt text between rounds, same number of repetitions, fresh sessions with no history, and the country and language written down on every run. Report citation frequency as valid runs citing your domain divided by total valid runs, and publish the denominator next to it. Keep the unfavorable runs too, because a log with the bad weeks deleted cannot show a trend. If you want the run automated rather than done by hand, the method is in how to program an agent to run your GEO maintenance.

Two second signals are worth pairing with it. Your own access logs identify Perplexity by user agent: per Perplexity's bot documentation, checked on 21 September 2026, PerplexityBot surfaces and links websites in Perplexity search results and is not used to crawl content for foundation models, while Perplexity-User visits a page to help answer a question a user just asked. Referral sessions from perplexity.ai in your analytics are a floor rather than a total, since referrer data gets lost along the way. Search Console's generative AI performance report, announced for worldwide rollout on 31 August 2026, reports AI Overviews and AI Mode impressions only, so it tells you nothing about Perplexity.

Be honest about the ceiling on all of this. Mention frequency inside your own battery is not share of voice and not exposure among real users, it is what one prompt set returned on one date. Third-party trackers run the same prompts for you, which is convenient and can be worth paying for, but their output is a sample as well, and a sample presented as a measurement is the same mistake either way. Keep "we were cited" separate from "we were described correctly", because a confident wrong description is its own problem and usually the more urgent one.

Start with the only first party data in the category, which is free. Search Console's generative AI performance report, announced for worldwide rollout on 31 August 2026, covers impressions from AI Overviews and AI Mode grouped by page, country, date and device. Checked against Google's own help page on 22 September 2026, it reports impressions only: no clicks, no average position, no query dimension and no split between AI Overviews and AI Mode. That is also the ceiling on what anyone can sell you about Google, so a dashboard showing your ranking inside AI Overviews is not reading Google data. It is running its own prompts.

The second category does exactly that, on a schedule, and records whether you were mentioned and which sources were cited. Useful, and worth paying for if it saves you the work, but the output is a sample rather than share of voice no matter what the dashboard calls it. Judge these products on evidence handling instead of on the size of the score: does it store the full response text with the engine, date, country and number of repetitions, can you export the raw log, and does it keep the runs where you did not appear. A single AI visibility score with no prompt list and no denominator behind it is a number, not a measurement. The method underneath is the same one described in our AEO guide for ChatGPT and Perplexity, and running it by hand for two weeks tells you what you would be buying.

The third category is ordinary technical crawling, and it is the one most brands skip while shopping for trackers. Google's AI features documentation, checked on 22 September 2026, says a page has to be indexed and eligible to be shown with a snippet, and that there are no additional technical requirements to appear in AI Overviews or AI Mode. Nothing on that list is exotic, and nothing on it requires a purchase. We do not publish a ranked vendor list on this page on purpose: this software category is barely a year old, the products change every quarter, and a ranking written today would be wrong before it was useful to you.

The verifiable part is real and worth doing. You can check that Googlebot, OAI-SearchBot, PerplexityBot and Claude-SearchBot receive a normal 200 response, that key pages are indexed, that structured data matches the visible text, and that each page answers a buyer's question in its opening lines. Each of those has a pass or fail state you can record with a date. Google's guidance on AI features, read on 6 October 2026, says a page must be indexed and eligible to show with a snippet to appear as a supporting link, so fixing indexing problems has a direct and documented effect on eligibility.

The outcome part is where honesty matters. Search Console's generative AI report shows impressions only, grouped by page, country, date and device, with no queries and no clicks, and Google includes AI feature traffic inside the normal Web totals. ChatGPT, Perplexity and Claude publish no comparable report for site owners. As a result, a claim such as a percentage rise in AI citations usually rests on a tool's own sample of prompts, which is useful as a signal but not proof of business impact. Ask how the prompts were chosen, how often they were run and what the baseline was.

A fair way to judge AI SEO for your business is a dated before and after. Record a prompt log and Search Console impressions for four weeks, make the access, indexing and content changes, then repeat the same log with the same prompts. If the brand starts appearing and inquiries mention AI tools, that is evidence. If nothing moves after a reasonable period, the limit is more likely demand or authority than formatting. Our guide to the Search Console generative AI report explains what that data can and cannot show.

The question list matters more than the tool. Write questions in the buyer's own words, mixing category questions such as the best provider of a service in a city, problem questions that describe a pain without naming a solution, and comparison questions that name you against a competitor. Avoid questions that include your brand name unless you are testing what the engines say about you, because a branded question measures reputation, while discovery is what you are trying to test. Ten to twenty questions is enough for a small business. Keep the wording fixed between rounds, since small changes in phrasing can change the answer and make the comparison meaningless.

Record results in a simple sheet with one row per question, platform and date. Useful columns are whether the brand was named, its position in the answer, the URL cited, the competitors named and a short note on accuracy. Run each question in a logged-out or fresh session where possible, because answers can reflect personal history. AI answers vary between runs, so a single test is an observation and not a baseline; three rounds over three weeks give a steadier picture. Paid trackers automate the same process at scale, but they still sample prompts, and the logic is the same.

Add the one official number Google provides. Search Console's generative AI report shows impressions from AI Overviews and AI Mode by page, country, date and device, which confirms whether Google is showing your pages in those features at all, and lets you see which pages carry that exposure. Pair that with your server logs, which record every crawler visit with a date: if OAI-SearchBot, PerplexityBot or Claude-SearchBot never visit, absence from those engines is an access problem first. Our GEO audit checklist lists the access checks to run alongside the prompt log.

Google's position, stated in its guidance on AI features and read on 6 October 2026, is that when people click from search results pages with AI Overviews, those clicks are higher quality, meaning users are more likely to spend more time on the site. That is Google's own claim about its product, so treat it as a declaration and not as independent evidence. The same page explains that traffic from AI features is counted inside the normal Web search type in Search Console, so there is no separate AI traffic line to compare against the rest.

The practical test is on your own property. In the Search Console performance report, compare clicks, impressions and click through rate for the same pages and the same months year over year, and look for pages where impressions held steady while clicks fell. Informational pages that answer a simple question are the most exposed, because the overview can satisfy the searcher before any click. Pages where the user needs to act, such as booking, buying, comparing prices or contacting a provider, tend to keep their reason for a visit. The generative AI report adds impressions from AI Overviews and AI Mode, which shows where your pages already appear inside those answers.

If clicks have fallen, the response is not to block Google's features, since snippet controls such as nosnippet also limit how a page appears in normal results. A better route is to make the pages that appear in overviews worth the click, with the detail, tools, prices or next step a summary cannot provide, and to strengthen pages tied to action. Whatever you change, record the date, so the next comparison in Search Console can show whether it helped. Our guide to the Search Console generative AI report explains how to set a baseline from impressions.

They are separate user agents. Allowing your content to be found and cited does not oblige you to allow it to be used for training, and the reverse is also true.

Decide on each one deliberately in your robots.txt rather than pasting a template from a blog post. And a reminder that gets skipped: robots.txt is a crawling instruction, not a security control. It protects nothing. If something must not be public, it needs authentication.

FAQPage structured data still earns its place, because it makes a question and its answer a clean, machine-readable pair that is easy for a retrieval system to lift. That is extractability, and it is worth having.

It is not a rich snippet any more, and anyone still selling it as one is working from a deck they have not updated in over a year.

Entity and structured data corrections propagate fastest, because they only need a recrawl to take effect. Cleaning up a domain that is carrying old junk pages or a confused identity often produces the first visible change.

Earning citations in competitive commercial categories takes longer, because it depends on third-party sources the engines already trust, and that trust cannot be bought outright. Any timeline shorter than this that comes with a guarantee is a sales device.

Gmail's sender guidelines set the baseline for any marketing automation program that emails consumers. Since 1 February 2024, anyone sending 5,000 or more messages a day to personal Gmail accounts must authenticate with SPF, DKIM and DMARC, align the From domain with SPF or DKIM, support one-click unsubscribe on marketing mail and keep the spam rate reported in Postmaster Tools below 0.30%, according to Google's own sender guidelines help page. Microsoft applied similar rules to Outlook.com: from 5 May 2025, high-volume domains without SPF, DKIM and a DMARC policy of at least p=none are routed to Junk, as announced on the Microsoft Defender for Office 365 blog. No automation platform can fix missing DNS records, so sender authentication is the first item to audit before a new sequence goes live.

Apple's Mail Privacy Protection changed what an email open means for automation triggers. Apple's support documentation says the feature hides the recipient's IP address, downloads remote content privately in the background when a message arrives rather than when it is viewed, and prevents senders from seeing whether the message was opened. A tracking pixel can therefore fire for an Apple Mail user who never read the email. Any automation that branches on an open, such as a follow-up email, a lead score increase or a sales alert, inherits that noise. Clicks, replies, form submissions, booked meetings and CRM stage changes are actions a person actually took, which makes them the more reliable triggers to build workflows on in 2026.

AI inside a marketing automation stack mostly handles drafting, classification and routing: writing message variants, tagging inbound replies by intent and summarizing a lead's history before a sales call. Those AI steps are cheap to run and easy to check against the source data. The decisions that still need a person are what is true about the business, which offer goes to which segment, and when a sequence should stop contacting someone. Results depend on list quality, offer and sending history, so any benchmark quoted without those three details says little about a specific account. A practical walkthrough of how these workflows fit together is in our marketing automation guide for 2026.

Most of the work in AI creative production happens before any frame is generated. A strategist writes the brief from a positioning decision and an audience insight, then a script and a shot list fix what each frame has to communicate. The next stage locks references: character sheets for faces and wardrobe, product renders, locations and a lighting look. Shots are then produced image-to-video from those references, because keeping the same face, logo or product across shots is the hardest thing to get from a text prompt alone. Selection is where the hours go: every generation is checked against the shot list, and many are discarded for artifacts, off-brand details or broken hands and text. The surviving clips are cut, graded, scored and captioned like any conventional spot, which is how our AI creative production process runs.

Ownership is the part of AI creative production most brands skip. The US Copyright Office concluded in Part 2 of its Copyright and Artificial Intelligence report, published January 29, 2025, that the mere provision of prompts does not make an AI output copyrightable, and that protection depends on a human determining sufficient expressive elements, for example through creative arrangement or modification of the output. The Office also said that using AI to assist creation, or including AI material inside a larger human work, does not by itself bar protection. In practice, the human-directed stages of script, selection, edit, compositing and sound are where that authorship sits, so the brief, reference sheets and edit files are worth keeping as a record. Anything that will run as paid media should still get a rights review from counsel.

Disclosure rules also shape what gets produced. YouTube requires creators to disclose realistic altered or synthetic content, such as footage that makes a real person appear to say or do something they did not, alters footage of a real event or place, or generates a realistic scene that did not occur. YouTube exempts clearly unrealistic content and minor edits such as color adjustment, beauty filters or upscaling, and treats AI help with scripts or thumbnails as production assistance that needs no label. Other platforms apply their own labeling rules, so the delivery checklist for an AI spot needs a disclosure decision per platform, made before launch rather than after a takedown notice. A production that plans for disclosure can use generated footage freely; what it cannot do is present a synthetic scene as a record of something that happened.

Google's Ads Help page on ads and AI Overviews says text and Shopping ads from existing Search, Shopping and Performance Max campaigns can show above, below or within AI Overviews, provided the ad is relevant to both the query and the Overview. Placement within AI Overviews is limited to English in 12 countries including the US, sensitive verticals such as finance and healthcare are excluded, advertisers cannot target a position inside the Overview, and there is no segmented reporting for these placements. Google's Search Central guidance on AI features is a separate document and states that there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible to show in Search with a snippet.

OpenAI's help center describes ChatGPT Ads as a separate paid format: ads appear below the end of a response, labelled as sponsored and visually separate from the answer. Ads are being rolled out gradually to eligible Free and Go users in select regions, and are not shown on Plus, Pro or Business plans. Advertiser context hints guide matching but are not exact-match keywords and do not guarantee delivery in specific conversations, and buying is by views, clicks or conversions. We read this in OpenAI's help center excerpts on 29 September 2026, and rollout details change often, so check the current pages before budgeting.

The practical split for a brand is simple. Ads are paid media, judged on clicks and conversions, and they can be switched on or off. Citation is earned: pages that are indexed, useful and clear about who the business is, plus crawler access that matches your goals. Any vendor selling a guaranteed citation in ChatGPT or AI Overviews is describing something neither platform documents. The GEO audit checklist covers the earned side, and no plan can promise a result there either.

Start a GEO audit with the baseline and the access check. The baseline is a fixed list of buyer questions, run in fresh sessions on ChatGPT, Perplexity and Google, with each answer saved and dated, plus the impressions shown in the Search Console generative AI report. Access is tested by requesting your pages with each crawler user agent and confirming a 200 response, because a server or firewall can rate limit a bot that robots.txt allows. Check the OpenAI bots separately: according to OpenAI's crawler documentation, read on 2 October 2026, OAI-SearchBot surfaces sites in ChatGPT search, GPTBot collects training data, and a robots.txt change takes about 24 hours to apply.

The second block of a GEO audit is eligibility and hygiene. Google's guide to AI features, updated 10 December 2025, says a page only has to be indexed and eligible to show with a snippet to appear in AI Overviews or AI Mode, with no additional technical requirements. The checklist therefore looks for noindex, nosnippet, data-nosnippet and max-snippet rules applied by mistake, and for the generative AI setting in Search Console. Hygiene means finding pages that speak for your brand without your knowledge: leftover CMS demo posts, old feeds, a forgotten sitemap, or spam left behind by a past hack. On our own domain, the audit found 28 theme demo posts served under the Cipion name.

The final block of a GEO audit is entity and content. Entity checks confirm that your Organization markup is one connected graph with sameAs links to profiles you control, that structured data matches the visible text on each page as Google recommends, and that your name, services and location read the same everywhere. Content checks ask whether each priority question has a page that answers it in the first 40 to 60 words, with the supporting fact and its source in the same paragraph. A checklist tells you what to fix and in what order. Citation can follow from that work over weeks, on a timeline no audit controls. For a scored version with 14 items, see our GEO audit checklist.

AI creative marketing differs from AI creative production in where the work is aimed. Production is the pipeline that makes a finished spot, while marketing is the loop that decides which message, format and audience combination earns the budget. Generative models make that loop cheaper: a team can draft ten headline angles, render product shots in several settings and cut one video into lengths for each placement in an afternoon. The models do not know which angle fits the positioning, which audience segment is worth paying for, or which variant will hold up once real spend and real comments arrive. A strategist still chooses the hypotheses worth testing, and a media buyer still reads the results against business numbers rather than platform metrics alone. Our AI creative service is built around that split.

Claims and endorsements are the area where AI creative marketing most needs a person with the authority to say no. Generative tools write fluent copy that can include product claims, statistics or customer quotes nobody verified. The Federal Trade Commission announced a final rule on August 14, 2024 that bans fake consumer reviews and testimonials, explicitly including reviews that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews, and that lets the agency seek civil penalties against knowing violators. A generated testimonial in an ad or on a landing page sits inside that risk. Every claim, number and quote a model drafts therefore needs a named reviewer who can trace it to a source the business can produce, and regulated categories should add a review by counsel before launch.

Search visibility follows the same logic for AI-assisted content on a brand's own site. Google Search Central published its guidance on AI-generated content on February 8, 2023, stating that appropriate use of AI or automation is not against its guidelines, while using automation, including AI, to generate content with the primary purpose of manipulating search rankings violates its spam policies. The same guidance suggests adding AI or automation disclosures where a reader might reasonably ask how a piece was created, and says that giving AI an author byline is probably not the best approach. A team using AI to draft blog posts, landing pages or product descriptions still needs an editor who checks facts, adds first-hand experience and decides whether each page deserves to exist at all.

Google's own guidance on AI features, last updated on 10 December 2025 and read on 6 October 2026, states that there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary. A page has to be indexed and eligible to show with a snippet, the same bar as a normal result. For Google, then, the sceptics are right: the foundation of GEO is crawlable, indexable, useful pages. Anyone selling a secret Google AI trick is selling something Google says does not exist, and the same guidance adds that no special schema.org markup or AI text file is needed.

The picture changes outside Google. ChatGPT, Perplexity and Claude run their own crawlers, OAI-SearchBot, PerplexityBot and Claude-SearchBot, and each one has to be allowed separately in robots.txt and served correctly by the server. A site can rank well on Google and still be closed to one of those agents. Writing also matters more than it used to, because an answer engine lifts a passage and not a whole page, so a paragraph that states the claim, its context and its source together is easier to reuse than marketing copy spread across a page.

Measurement is the third difference and the hardest one. Search Console reports clicks and positions for classic results, but its generative AI report shows impressions only, by page, country, date and device, and the other engines publish no equivalent report for site owners. GEO therefore relies on dated logs of repeated prompts, which are slower and noisier than a rank tracker. Calling GEO a new name for SEO is fair for the Google part, and incomplete for everything else. Our GEO audit checklist shows where the two overlap.

Google's guidance on AI features, read on 6 October 2026, says there is no special schema.org structured data you need to add to appear in AI Overviews or AI Mode. Its general structured data policies, last updated on 10 July 2026, add that structured data enables a feature to be present but does not guarantee that it will be present. Taken together, those two statements rule out schema as a lever for AI citations in Google. A page with perfect markup and weak content is still a weak source, and a clear, well written page without markup can still be quoted.

Schema is still worth having for a different reason: it removes ambiguity about entities. Organization markup with a stable identifier, the correct legal name, address, phone number and sameAs links to real profiles helps any system connect your website to the same business described elsewhere. Product, service and article markup describe what a page is about in a form machines parse reliably. The value is consistency, which supports every engine, rather than a ranking boost in one of them. That work belongs in a technical review and should be described as such.

The rule that matters most is the one most often broken. Google's policies say not to mark up content that is not visible to readers of the page, and markup that describes something the page does not show can stop the feature from working. FAQPage schema is a common example: the questions and answers in the markup must appear on the page, and Google retired FAQ rich results on 7 May 2026, so the markup now helps readability only. We describe the fields that fail without warning in our GEO audit checklist.

Cipion prices per engagement rather than from a public rate card, because the honest number depends on the size of the site, the state of its structured data and how contested the category is. Quoting before seeing those is guessing with extra steps.

The first conversation and the AI Visibility Snapshot are free, and you get the findings whether or not you hire us. The full breakdown of what drives the cost is here.

The last one is the most revealing. Any agency guaranteeing a position in AI answers is selling something no engine currently offers. There is no paid placement in an AI Overview, no submission form for ChatGPT, and no mechanism to buy a citation.

The baseline question matters nearly as much. If nobody measured your citation share before starting, nobody can prove the work did anything afterwards. More on evaluating agencies here.

Evidence you can check starts with the agency's own website, because an agency selling AI visibility is running the same experiment on itself. Read its robots.txt. OpenAI documents OAI-SearchBot as the crawler behind ChatGPT search and GPTBot as a separate crawler for model training, so an agency that blocks the first, or treats the two as one, has not applied the documentation it sells. Then check whether its structured data parses and whether its claims carry dates and named sources. Our GEO audit checklist lists what to look at.

Ownership is the criterion most often skipped. Before signing, confirm in writing that prompt logs, dashboards, analytics properties, ad accounts, content files and any automations are created under your accounts, or handed over when the engagement ends. Work you cannot take with you is rented, and it makes switching agencies later far more expensive than it needs to be.

The two answers above cover what the work tends to cost and what to ask once you are on a call. These criteria are the filter that decides who gets that call.

The tasks ChatGPT handles well are repetitive and easy to check. It can turn a list of customer questions into page outlines, rewrite a long paragraph so the answer comes first, suggest title tags within a length limit, draft schema from details you provide and compare a page against an audit list you paste in. Each output is fast to review, so the time saved is real. Pairing it with a script that fetches your live pages gives better results than asking it about your site from memory, because a model working from its training data may describe an old version of the site or invent pages.

The risks are accuracy and scale. A model will write confident claims about prices, results or credentials that you never gave it, and publishing those claims is your liability, not the tool's. Google's spam policies also include scaled content abuse, which covers producing many pages mainly to manipulate rankings, whatever tool generates them. Publishing dozens of near identical AI pages on a recently cleaned domain is the pattern that policy describes. Every factual statement should come from a source you can name, and every page should exist because a buyer needs it.

The part that stays human is judgment. Deciding which services to promote, which pages should exist at all, how to answer a question honestly when the evidence is thin, and when to stop publishing are business decisions that need context a model does not have. A workable setup is an agent that runs checks and drafts changes, with a person approving what goes live. We describe that setup, with its guardrails, in how to program an AI agent to run your GEO maintenance.