How AI Overviews and ChatGPT Actually Pick Which Contractor to Recommend

A roofing company outside Fort Worth showed us a proposal last spring. Page four promised “AI citation optimization” and listed the deliverables: an llms.txt file, AI-specific schema markup, and twelve blog posts a month. The retainer was $3,400. Nothing in it was tied to a single published source, and the one number on the page — a claim about AI referral growth — traced back to another agency’s blog post, which traced back to nothing.

The explanation most contractors get is that AI engines are a new search engine with new rules, so you need a new service to win them. That framing is what sells the retainer. It is also mostly wrong, and the people best positioned to know say so out loud — Seer Interactive, an agency running one of the largest published datasets on this, states plainly that “no one is a GEO expert.”

The real mechanism is knowable, because Google and OpenAI have both documented parts of it and Ahrefs has published the largest correlation study anyone has run. What comes out of those three sources is less exciting than the pitch and far more useful: AI answers are assembled in two stages, the inputs that predict visibility are mostly reputation signals you already understand, and the input most agencies sell — volume of published pages — has close to no measured relationship with the outcome at all.

This piece walks the mechanism that decides how AI picks contractors end to end: what query fan-out is and why it changes which pages get seen, what the overlap data says about the relationship between Google rankings and AI citations, what the correlation study found and what it explicitly doesn’t prove, and what all of that means for where a Texas HVAC or plumbing company should actually put its money.

What’s in this guide

1. The 12% problem
2. Query fan-out, in Google’s own words
3. Being retrieved and being cited are two different fights
4. What the correlation data actually shows
5. What Google says you need — and says you don’t
6. What E-E-A-T means when the reader is a machine
7. Where local trades work differently
8. What this means for your actual priorities
9. Questions contractors actually ask

The 12% problem

Ahrefs ran 15,000 long-tail queries through ChatGPT, Gemini and Copilot in early July 2025 using its Brand Radar tool, then checked how many of the URLs those engines cited also appeared in Google’s top ten for the same query.

The answer was 12%.

Roughly 80% of the URLs cited in AI answers didn’t rank anywhere in Google for the query that triggered them. Per engine, the overlap ranged from Perplexity at 28.6% down to Copilot at 6.1%.

How much AI citation overlaps with Google’s top 10

Share of AI-cited URLs that also rank in Google’s top ten for the same query. 12% overall across all engines tested.

Take that seriously and it reframes the whole service. Ranking first in Google does not automatically get you cited, and being cited does not require ranking first. They are overlapping but largely separate retrieval problems. It also means the honest answer to “will my SEO carry over to AI?” is: partially, and less than your vendor implies.

The limitation matters as much as the headline. Long-tail informational queries are exactly the kind of question where an AI engine goes hunting for niche sources. A homeowner typing a commercial local query is a different retrieval shape, and nobody has published a clean local-trades cut of this. Anyone quoting the 12% figure at you as if it were measured on “best AC repair in Austin” is overstating what the study did.

Query fan-out, in Google’s own words

The mechanism behind that low overlap has a name, and Google describes it directly. In AI Mode, rather than running your query once, Google uses a technique it calls query fan-out — “issuing multiple related searches across subtopics and data sources.” The stated consequence is that AI Mode surfaces “a wider and more diverse set of helpful links.”

Sit with what that does to your competitive position. You have spent years trying to rank for one phrase. The system now decomposes a homeowner’s question into six or eight sub-questions you never targeted — what a repair typically costs, whether a permit is required, what warning signs precede a failure, which brands are reliable in a hot climate — and goes looking for a source for each.

That’s simultaneously bad news and the best news in this article. Bad, because your money keyword is no longer the only door. Good, because there are now eight doors, most of them long-tail, most of them unguarded, and most of them answerable by somebody who actually does the work for a living.

Being retrieved and being cited are two different fights

Almost every argument about AI optimization collapses because people conflate two stages that have different rules.

Stage one: retrieval

The engine has to find your page at all. This is where crawler access, indexation, and your presence in whatever search index the assistant queries decide everything. OpenAI’s publisher documentation states that “any public website can appear in ChatGPT search,” and that its search crawler is OAI-SearchBot — separate from GPTBot, which handles training. If you’re blocked here, nothing downstream matters. We walk that check step by step in the ChatGPT visibility diagnostic.

Stage two: selection and citation

Given a pool of retrieved sources, the model decides which to lean on and which to name. This is where the Princeton and IIT Delhi research on generative engine optimization applies — and where its findings are routinely misrepresented, because that study measured what happens to a source already inside the retrieved set. It says nothing about how a contractor gets into that set in the first place. We unpack that distinction, and the dollar figures attached to it, in what AI search optimization actually costs.

Most contractors are losing at stage one and being sold stage-two tactics. That’s the single most expensive mistake in this category.

What the correlation data actually shows

Ahrefs’ second study is the largest thing published on this question: 75,000 domains with a Domain Rating above 40, each with a highest-volume keyword of at least 800 monthly searches, measured against millions of AI responses across ChatGPT, AI Mode and AI Overviews. They ran Spearman correlations between AI brand visibility and a set of measurable brand signals.

What correlates with being visible in AI answers

Spearman correlation with AI brand visibility across ChatGPT, AI Mode and AI Overviews. 75,000 domains.

Two things on that chart deserve to change how you buy this service.

YouTube mentions correlated most strongly of anything tested, on every engine. Not your website. Not your domain authority. Video, and being talked about in it.

The number of pages on your site sat at the bottom, around 0.19 — effectively no relationship. That is the deliverable most AI optimization retainers are actually made of. The largest published dataset on the question says it barely moves.

Ahrefs is careful about what this means, and we’ll repeat their caution rather than bury it:

Correlation isn’t causation.
Ahrefs — AI brand visibility correlations

Big brands get mentioned on YouTube and get cited by AI because they’re big brands. The correlation doesn’t prove that filming videos causes citations. It also doesn’t disprove it, and the direction is consistent enough across three independent engines to take seriously as a planning input. What it rules out fairly firmly is the idea that pumping out pages is the lever.

What Google says you need — and says you don’t

Google publishes documentation on AI features and your website. It is short, it is specific, and it eliminates an entire category of product being sold to contractors right now.

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.
Google Search Central — AI features and your website

And, more pointedly:

You don’t need to create new machine readable files, AI text files, or markup to appear in these features. There’s also no special schema.org structured data that you need to add.
Google Search Central — AI features and your website

Common AI optimization deliverables against the documentation

Being sold as essential	What the primary source says	Verdict
An llms.txt file	Google: you don’t need to create “new machine readable files, AI text files, or markup” to appear in AI features.	Not required
“AI schema” or GEO-specific structured data	Google: “There’s also no special schema.org structured data that you need to add.”	Not required
Paid placement or submission to ChatGPT	OpenAI: “Any public website can appear in ChatGPT search.” Ads, tested in the US from February 9 2026, are labelled sponsored and “do not influence ChatGPT’s answers.”	Doesn’t exist
High-volume blog publishing	Ahrefs, 75,000 domains: number of pages on site correlates at roughly 0.19 with AI visibility — the weakest signal tested.	Weak evidence
Crawler access for OAI-SearchBot	OpenAI documents OAI-SearchBot as the search crawler; blocking it removes you from the pool.	Genuinely required
Third-party mentions of your brand	Ahrefs: branded web mentions 0.656–0.709, YouTube mentions ~0.737 — the strongest correlations measured.	Best-evidenced lever

What E-E-A-T means when the reader is a machine

Google’s quality framework — experience, expertise, authoritativeness, trustworthiness — was written for human raters assessing whether content deserves to rank. It isn’t a switch in an algorithm and it isn’t an AI-specific requirement. But it describes, better than anything else Google publishes, the qualities that make a page worth quoting.

For a contractor, the first E is the one you already have and almost never use. Experience is the thing an agency copywriter in another state cannot fabricate: what a 4-ton unit actually costs installed in your market this quarter, which neighborhoods have the cast-iron drain lines, what the inspector in your county flags every single time, how long a hail claim actually takes to settle in Dallas–Fort Worth.

Most contractors stop at “we’ve been in business 22 years.” That is not expertise, it’s tenure. Expertise on the page looks like a number you had to earn: we replaced 61 condensers in Bexar County last summer and 14 of them failed for the same reason. That’s a sentence a generative engine can lift, attribute and build an answer around — and one no competitor can copy without doing the work.

It’s also why generic city pages don’t get cited. A page that says “we proudly serve Katy and the surrounding areas” contains nothing quotable. The fix isn’t more pages; it’s pages with facts only you could know, which is the same argument we make for winning cities you can’t reach in the map pack.

Where local trades work differently

Every study cited so far treats the open web. Local commercial queries behave differently in one important respect: the assistants lean heavily on directories.

In Mecha’s tracking of plumbing and HVAC recommendations across ten US markets — vendor research, small sample, but with methodology stated — ChatGPT cited sources roughly 40% of the time, and those sources were frequently Yelp or Angi rather than contractor websites. In three of the ten markets, Perplexity recommended a company that had no website at all, drawing entirely on Yelp and BBB data.

Treat those percentages as indicative rather than precise. But the behaviour is consistent with what the open-web correlation data suggests: what’s said about you elsewhere carries more weight than what you say about yourself. For a local contractor, “elsewhere” is mostly reviews and directory profiles — which is why we treat Google Business Profile accuracy and steady review flow as AI work, not just local SEO work.

Scale matters too. Whitespark’s guide to Google’s AI Mode for local businesses reports AI Overviews appearing on roughly 68% of local searches in its Q1 2025 data, and cites GatherUp Q3 2025 research finding 48% of people had used a conversational AI tool while researching a local business — with 67% not rigorously fact-checking the sources those tools cited. We reached Whitespark’s guide, not the GatherUp original, so we’re citing the intermediary and saying so.

What this means for your actual priorities

Where a contractor’s effort actually belongs

Priority	What it is	Why it’s ranked here	Who does it
1	Crawler and indexation access	Binary gate. Documented by OpenAI. Costs nothing to fix.	Whoever controls your site files
2	Steady review flow and accurate profiles	Third-party text about your brand at volume; directories are what the assistants actually cite in local trades.	Your office, weekly
3	Being mentioned off your own site	Strongest measured correlation in a 75,000-domain study.	You, via suppliers, associations, local press
4	Video of real work	YouTube mentions correlated ~0.737 — the top signal on all three engines tested.	A phone and a crew lead
5	A small number of genuinely specific pages	Feeds query fan-out. Quality, not count — page volume correlated at ~0.19.	Someone who interviews your techs
6	llms.txt, AI schema, GEO audits	Google states these aren’t required. Not harmful; just not the lever.	Nobody, until something changes

Questions contractors actually ask

What to do with this

The mechanism isn’t mysterious. Engines decompose a question, retrieve a pool of sources from an index you may or may not be in, and then choose which of those sources to name. You can influence all three stages, but only one of them is fixed this week and none of them is fixed by a file upload.

So start where the evidence is strongest and the cost is lowest. Confirm you’re crawlable. Keep reviews and directory profiles current, because in local trades those are frequently the sources being read. Get mentioned somewhere that isn’t your own website — a supplier’s contractor page, a trade association, a local news story, a YouTube video of a job you did last Tuesday.

Then write a small number of pages that contain something only you could know, and let fan-out find them. That’s the whole strategy. It looks a lot like good local SEO and good website content, because according to the best data anyone has published, that’s what it is.