6 min read
Open ChatGPT. Type: who is the best [your category] in [your city]? Read the answer.
That is the whole diagnostic. It takes about ten minutes to do properly, it costs nothing, and it tells you something your rank tracker cannot. This piece covers how to run it so the result means something, what the answer is actually made of, and what to do if your name is not in it.
The ten-minute test
Do not ask an AI whether it recommends you. Ask it the question a customer would ask. Those are different questions and they get different answers.
Run three types of prompt across three surfaces — ChatGPT, Gemini, and Google AI Mode. Use a logged-out or temporary session so your own history does not tilt the result.
- The direct ask. “Who is the best [category] in [city]?” This is the recommendation question. It is the one where being unnamed costs you the most.
- The qualified ask. “I need a [category] in [neighborhood] that handles [specific situation]. Who should I call?” Buyers are specific. Generic prompts flatter incumbents; specific prompts are where a smaller business can actually win.
- The comparison ask. “Compare [you] and [your closest competitor].” This tells you what the model believes about you — which is often more useful than whether it names you.
Record three things for each: whether you appear, who does appear, and what sources are cited. The citations are the most valuable part of the exercise. They are a map of which pages the model trusts for your category and city, and most of them will not be anybody’s homepage.
Why your Google ranking did not save you
The reasonable assumption is that ranking well on Google means getting named by AI. That assumption is getting weaker fast.
Ahrefs analysed 863,000 keywords and roughly four million AI Overview URLs in March 2026. Of the pages cited inside AI Overviews, 38% ranked in Google’s top ten. Another 31.2% ranked between 11 and 100. And 31% came from pages ranking beyond position 100 — or not ranking at all. In a comparable study in July 2025, the top-ten figure was 76%.
The honest caveat: Ahrefs attributes part of that drop to improved parsing on their side rather than pure change in Google’s behaviour, so the 76-to-38 swing is not a clean year-over-year comparison. Take the direction seriously and the precise magnitude with some salt.
Source: Ahrefs, March 2026, via Search Engine Journal.
The second number matters more for the business case. SparkToro, using Similarweb panel data for US Google searches from January through April 2026, found that 68% of searches ended without a click — up from 60.45% in 2024. SparkToro is explicit that long-run comparisons across different data providers are not directly equivalent.
Source: SparkToro / Similarweb, 2026, via Search Engine Land.
Put those together and the shape of the problem is clear. Most searches now end in an answer rather than a visit, and the sources behind that answer are only loosely related to who ranks. Ranking is still worth having. It is no longer sufficient on its own.
31%of pages cited inside Google’s AI Overviews rank beyond position 100 — or do not rank anywhere at all. Ahrefs, 863,000 keywords and roughly four million AI Overview URLs, March 2026.
What the answer is actually made of
When a model names a business, it is not expressing an opinion. It is reconstructing a consensus from whatever it can retrieve. Four inputs do most of the work.
1. Entity consistency
Your name, address, phone number, and category have to agree across every place they appear — your site, Google Business Profile, Apple Business Connect, Bing Places, the directories, your social profiles. Disagreement does not just look sloppy. It makes you harder to resolve as a single entity, and an entity a system cannot resolve confidently is an entity it will not name.
2. Third-party corroboration
Models weight what other sources say about you far more heavily than what you say about yourself. A local news mention, an industry directory, a “best of” roundup, a podcast appearance, a review body with real volume behind it — these are the raw material of a recommendation. A beautifully written About page is not.
3. Structured data
Schema markup does not make you rank. It makes you machine-readable: what you are, where you are, what you charge, when you are open, what you do and do not handle. That is the difference between a page a system can quote and a page it has to guess about.
4. Content that answers the actual question
Pages built to rank are often written for a keyword. Pages that get cited answer a question completely, in a way that can be lifted out and reused. If a paragraph on your site is the cleanest available answer to a question buyers ask, it will get pulled. If your answer is spread across four paragraphs of throat-clearing, something more compact will get pulled instead.
Run the test logged out, or in a temporary session. A logged-in window is tilted by your own search history, so the answer you see is not the answer your customer sees.
What to do if you are not in the answer
In order, because the order matters more than the list.
- Fix consistency first. It is the cheapest work with the most direct effect, and every later step compounds off it. Audit every listing you can find and make them agree exactly — including the abbreviations.
- Claim and complete the profiles you own. Google Business Profile, Apple Business Connect, Bing Places. Fill in every field, including the ones that feel redundant. Categories and service areas especially.
- Mark up the pages. Organization or LocalBusiness, Service, FAQPage where you genuinely have questions and answers, and Person for anyone whose name carries weight in your field.
- Build corroboration deliberately. Pick five sources that already rank or get cited for your category and city, and work out what it would take to appear on each. This is the slowest part and the part with the longest tail.
- Rewrite the top of your key pages to be quotable. Lead with the answer. Put the specifics — prices, coverage, exclusions, timelines — in text rather than in an image or a PDF.
- Re-run the test monthly and log it. Without a baseline you cannot tell progress from noise, and AI answers are noisy. One run is an anecdote. Six months of runs is a measurement.
Two things worth being skeptical about
Anyone guaranteeing you a spot in an AI answer. Nobody controls these systems, they change without notice, and the same prompt can return different answers to different people on the same day. What can honestly be committed to is a measured baseline, a defined scope of work, and reporting on the leading indicators.
Anyone selling this as a replacement for SEO. It is not a replacement. Classic ranking, AI citation, and outright recommendation are three jobs that share most of their inputs and diverge at the edges. Running them as one program is cheaper than running them as three, and running only one of them is how you end up invisible in the surface that is growing.
If you would rather not run the test yourself, that is what the free AI Visibility Snapshot is: ten real buying questions for your category and city, run across ChatGPT, Claude and Gemini, with the results and the citation sources written up. No call required, and if it shows you are already being recommended, you should keep your money.
More on how the three layers fit together in what we do.