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.

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.

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.

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.