What Actually Gets You Cited by AI: 9 Lessons From a Year of Tracking AI Visibility

The short version: we run AI-visibility tracking across a set of local-service and B2B sites, measuring whether ChatGPT, Perplexity, Google AI Overviews, Gemini, and Claude name and cite them. The lessons below are what the data actually showed, not what the checklists promise. The headline: a clean AEO-readiness score does not buy you citations, the biggest lever is mostly off your page, and the five engines disagree with each other far more than anyone selling you a single playbook will admit.

A note on how we report this. The first-party numbers here are aggregates across the sites we track. We do not attach percentages to named clients, we do not publish our own weak spots as if they were yours, and we never invent a figure to make a point land. Where a stat comes from someone else's research, we link it. That discipline is the whole job, and lesson 9 is about why.

1. A high AEO-readiness score is not AI visibility

This is the finding that reorganized how we work. We audited two sites on the usual technical AEO-readiness criteria: answer-first structure, schema, crawlability, clean HTML, fast rendering. One scored 67 out of 100. The other scored 59.

The 67 earned almost no AI citations across the engines we checked. The 59 was named on close to half the prompts we tested.

Readiness measures whether an engine can read and parse you. Visibility measures whether it chooses you. Those are different questions, and most AEO tooling only answers the first one. If you have paid for a readiness scan and come away thinking the score is your AI visibility, you have measured the wrong thing. We wrote a separate guide on how to actually check your AEO score and how to track AI visibility over time, because the gap between the two is where most of the real work sits.

2. The biggest lever is mostly off your page

Here is the uncomfortable part. The strongest predictor of whether an AI names you is often whether you were already in its training-data shortlist before it ran a single search.

Suganthan Mohanadasan published a teardown of ChatGPT naming brands before it searches: in 21 of 27 tested conversations, the first search query ChatGPT issued already contained brand names nobody typed. Brands that showed up in that first search were mentioned in the answer 68.9% of the time. Brands only fetched later sat at 2.1%. That is roughly a 33x gap, and the shortlist comes from what the model learned in training, not from what is on your page today.

Our own tracking lines up with this. In newer, less contested topics we get cited readily. In dense, established verticals where a few brands own the category in the training data, on-page work barely moves the needle. The levers that do move it are slow and external: digital PR, trusted-publication mentions, and genuine entity presence that feeds the next training cutoff. There is no on-page trick that shortcuts it. Anyone promising one has not measured it.

3. Geo-bound queries are winnable. Un-geotagged ones usually are not

This is the single most useful pattern we have measured, and it is almost never discussed.

When a query contains a place ("asbestos testing in Windsor Ontario", "pest control Scarborough"), the business can win the citation. When the same query is general and un-geotagged ("what is this regulation", "does this material contain X"), it mostly cannot. Across more than 470 un-geotagged informational checks in our tracking, we recorded zero citations. Every single one went to a canonical source like a government page, or to a consensus forum like Reddit, and no amount of page quality displaced them.

The honest caveat has to travel with that finding: geo-bound is necessary, not sufficient. Even inside the winnable class, only one of the sites we track converted geo-bound queries at a strong rate. The others, carrying near-identical pages, converted in the low single digits. So "put the city in the title and win" is the wrong takeaway. The right one is: un-geotagged informational queries are a near-closed door, and geo-bound queries are a door that opens, after which something else (authority, age, off-site presence) decides whether you walk through.

4. Being mentioned is not the same as being cited

When we started, we counted any appearance of a brand in an answer as a win. That was a mistake worth flagging, because the two outcomes have different levers.

A site can be mentioned in the prose of an answer (named in a sentence the model wrote) while never being cited as a source (the clickable URL the answer is built on). In our data these diverge constantly: a site named in informational answers was almost never used as the source for them. If your reporting lumps the two together, you will congratulate yourself for prose mentions while the citation, the thing that actually sends qualified traffic and compounds authority, goes to someone else. Measure them separately.

5. The five engines disagree more than they agree

There is no "optimize for AI search" in the singular. For the same set of queries in our tracking, citation rates differed by more than 80 points depending on which engine we asked. A topic we owned on one engine we were invisible on across another. That is not noise. It is structural, and it means any single aggregate "AI visibility" number is averaging over systems that do not behave alike.

A few of the engine-specific mechanics that are public and worth knowing:

• Perplexity never skips the web. Suganthan's Perplexity teardown describes an intent classifier with fixed thresholds; for local queries, actual businesses take the citations and "best X in city" listicles take none. If you run a local business, Perplexity is often your most winnable channel.
• Claude leans on Brave. The GSC Wizard Brave playbook cites two independent studies (Profound, March 2025, and MERJ, July 2026) putting 79% to 87% of Claude's citations inside Brave's top 10 results, versus around a third for Google's top 10. If you cannot measure Claude directly, your Brave ranking is the closest proxy you have.
• Google AI Overviews and ChatGPT are the hardest rooms. In our tracking these two, which also have the largest user bases, were the stingiest with citations for the sites we monitor. Winning them leans heavily on off-site signals, not on-page polish.

Pick your engine fights on purpose. Optimizing blindly for "AI" means optimizing for an average that describes none of them.

6. The technical table stakes still gate everything

None of the above matters if the crawler never gets a clean read of your page. Three failure modes we check first, all drawn from Suganthan's 15-point AI SEO playbook and verified on our own sites:

• Server-side render the content. If your words only appear after client-side JavaScript runs, most AI crawlers never see them. The content has to be in the HTML.
• Keep time-to-first-byte under about half a second. AI crawlers abandon slow fetches and move on to cite a faster competitor. In server logs the abandonment shows up as 499 errors.
• Do not block the Common Crawl bot. Suganthan flagged that 500,000-plus sites block CCBot, most of them by accident, because some CDNs and plugins add the block by default. Common Crawl is a major training input, so blocking it quietly removes you from the very shortlist that lesson 2 is about. Check your robots.txt and your CDN bot settings today.

These will not make you visible on their own. They are the price of being considered at all.

7. Your reviews are becoming the literal text of the answer

In August 2026, Yelp licensed its reviews, ratings, and photos to OpenAI. ChatGPT's local-business responses now carry a review-highlights field. As licensing deals like this fill in, the engines will paraphrase real review text directly into their answers.

That changes the stakes of your review corpus. An unanswered negative review is no longer just a bad look on a listing page a few people read; it becomes raw material an AI can fold into the answer a buyer sees first. The practical move is unglamorous and overdue: reply to every negative review, because your reply becomes part of the text the model reads, and keep the recent, substantive reviews coming.

8. Optimize for the searches the engine invents, not the one your customer typed

Modern AI search does not run one query. It fans a single prompt out into roughly ten related sub-searches, many of which have no measurable search volume, and assembles its answer from the union of those results. If your content only matches the exact phrase a person typed, you miss most of the fan.

This is important enough that we gave it its own piece: query fan-out, and how to get cited across the whole fan. The short version for here is that comprehensive, genuinely answer-first pages that cover a topic and its neighbours get pulled into more of the sub-searches than thin pages targeting one keyword. One strong page beats fifty thin ones, partly because engines rarely cite two pages from the same domain, and partly because the thin ones never match the sub-queries you did not think to write for.

9. The thing that protects all of it: we do not make numbers up

The reason this post can show you real figures is that we refuse to fabricate them. When we build first-hand-experience content and the real data is not in yet, we mark the gap and leave it empty rather than filling it with a plausible-sounding invention.

That is not only an ethics line, it is an AEO strategy. The engines are getting better at cross-referencing a claim against the sources around it, and a single detected fabrication can taint the trust the whole domain has built. A benchmarks page with one made-up statistic is worth less than no page at all, because it puts every other true thing you said under suspicion. The honest version is slower to write and it is the only version that compounds.

One encouraging note to end on. In a newer vertical where no shortlist is locked yet, our own published, answer-first content gets cited directly, homepage and blog alike, because the training data has not already handed the category to an incumbent. New and emerging topics are the most winnable ground in AI search right now. If your space is still forming, publishing the clearest, most honest, most comprehensive answer is the closest thing to a shortcut that actually exists.

What to do with this

• Separate readiness from visibility. Fix the technical table stakes in lesson 6, then measure citations, not scores. Check your AEO score, then track visibility over time.
• Invest in the off-page lever. Digital PR, trusted mentions, and entity presence are slow, but they are what moves the shortlist in contested verticals.
• Lead with geography where you serve a place, and stop expecting un-geotagged informational pages to earn citations they structurally cannot.
• Report per engine, and pick your fights. Local business? Prioritize Perplexity and your Brave ranking. Measure mentioned and cited separately.
• Treat reviews as answer copy, reply to the negatives, and keep them fresh.
• Write for the fan-out, one comprehensive page per topic, not fifty thin ones.

If you want the checklist version of the on-page side, we keep one current: the 8 things AI engines check before they cite you. And if you would rather we run the measurement and the off-page work for you, that is what our SEO and AEO service does.

FAQ

Does a good AEO-readiness score mean I will get cited by AI? No. Readiness measures whether an engine can crawl and parse your page. Visibility measures whether it chooses you as a source. In our tracking, a site scoring 67 out of 100 on readiness earned almost no citations, while a site scoring 59 was cited on close to half the prompts we tested. Fix readiness as table stakes, then measure citations separately.

What is the single biggest factor in whether AI recommends my business? Often it is whether your brand is already in the model's training-data shortlist before it searches. Research by Suganthan Mohanadasan found brands in the first search query are mentioned about 69% of the time versus about 2% for brands fetched later. That shortlist is built off your page, through mentions and entity presence, not by on-page tweaks alone.

Which AI engine is easiest for a local business to get cited by? In our tracking the engines differ by more than 80 points on the same queries, so there is no single answer. For local businesses, Perplexity is frequently the most winnable, since its local results favour actual businesses over "best of" listicles, and your Brave ranking is a useful proxy for Claude. Google AI Overviews and ChatGPT tend to be the hardest and lean most on off-site authority.

Why do my pages get mentioned by AI but not cited as the source? Being named in the prose of an answer and being used as the clickable source are different outcomes. Mentions come from your entity being known; citations come from your page being the best retrievable answer for one of the sub-searches the engine ran. If you want the citation, you need comprehensive, answer-first content that matches the query fan-out, plus the retrievability basics like server-side rendering and fast response times.

Will blocking AI crawlers protect my content or hurt me? For visibility it hurts you. Blocking the Common Crawl bot, which many sites do by accident through CDN defaults, removes you from a major training input and therefore from future model shortlists. If your goal is to be cited, let the AI crawlers through and keep your time-to-first-byte under about half a second so they do not abandon the fetch.

• $3,000 Growth CreditFree assessment · credit applied to onboarding. See if you qualify →

Want the optimization playbook, not just the platform overview? Get our free ChatGPT Ads cheat sheet — context hints, bid-floor tests, industry readiness, and the measurement stack from practitioners running real budgets.

Quick answers