Winning clients through ChatGPT: how a question becomes a shortlist

How a question becomes a shortlist

The question is specific before it has a name in it: "accountant in Sacramento who works with restaurants", "quiet hotel near the Denver convention center with parking", "emergency plumber in Mississauga that comes out at night". An AI assistant with web retrieval on does not answer that from memory. It turns the question into a few searches of its own, a step Google's documentation calls query fan-out, fetches a handful of the pages those searches return, and writes an answer from what those pages say. Each assistant runs its own version of that loop; what they share is that the pages are read for this question, at the time it is asked.

The shortlist is the businesses whose fetched pages match the constraints in the question. Not the best businesses, which the assistant cannot judge. The ones whose pages say, in sentences, that they do this thing in this place for this kind of buyer, and that give the assistant a name, an address and a phone number it can hand back.

We measured a narrow version of this. In our audit of 9,483 small business websites across 40 trades and 90 metros in the US, Canada, Europe and the UAE, run between late August and early September 2026, we put one buyer question per business to Claude with web retrieval on, 9,240 questions in all. 14.9% of the businesses were named; among the 5,928 questions the assistant answered with a differentiated list rather than a hedge, 17.7%. That is a Claude figure, one question each, one day, and it says nothing about ChatGPT, Gemini or Perplexity. What it does say is that most small businesses are not on the shortlist for their own question, and on inspection the ones that were tended to have the consistent identity and the answer-shaped content described below.

The four things your own site has to state

Identity, readability, answers and evidence. Identity is one name, one address, one phone number, repeated identically on every page and in every listing, so that a machine reading three pages concludes they describe one entity. Readability is the text being in the HTML a fetcher receives: not behind a bot wall, not rendered only by JavaScript, not in a PDF. Answers are the buyer's questions answered on the page in plain sentences: services by their names, the area served, prices or ranges where the trade quotes them, hours, credentials. Evidence is what someone other than you says: reviews quoted on the page, a rating with a count, named people.

The order matters. Answers written on a page the assistant cannot fetch are not read; evidence attached to a business the assistant cannot pin down is attached to nothing. Fix identity and readability first, then write the answers, then add the evidence. LocalBusiness schema, or the trade's specific type, carries the identity and the answers in machine-readable form, and Google's introduction to structured data is clear that it has to describe what a visitor can see.

One study bears on what to put in the answers themselves. Aggarwal and colleagues, in "GEO: Generative Engine Optimization" (KDD 2024), tested changes to source pages on a benchmark of ten thousand queries and found that adding citations, quotations and statistics raised a source's visibility in generated answers by up to about 40%. Their benchmark was not small business sites and their figure is not ours. The mechanism is the useful part: specific, attributable facts get quoted; adjectives do not.

How mentions off your site feed the answer

The assistant does not read only you. It reads the map listing, review sites, directories, the local paper, association member lists, supplier pages, a neighborhood forum. Each is a brand mention, and each one that carries your name in the same form as your site does helps the assistant conclude that the business it found on three pages is one business, with reviews here, a listing there and a site that states the facts.

Mentions do two jobs. They corroborate: a business named on its own site and nowhere else is a claim; a business named on its own site, a review site and the chamber of commerce list is an entity. And they carry facts your site does not, above all reviews, in words you did not write. What they cannot do is replace the site. In our probe, 99.4% of the names Claude gave were businesses and 0.6% were directories or aggregators, so the answer sends the buyer to you, and the page they land on had better say what the answer said.

The mentions worth earning are the ones a buyer in your trade would trust: the profile Google's Business Profile guidelines describe (real name, real address, categories, hours), the review sites your buyers already use, the association your trade belongs to, the local press when you have something to say. Keep the name, address and phone number identical in every one of them. A mention under a variant name is a mention of someone else.

How to measure whether it is working

With a fixed panel of questions, put to an assistant on a schedule, with every answer kept. Our method: twelve questions written the way the client's buyers type them, set once per trade and location and then left alone so that readings compare; every question put to one assistant, Claude, with live web search on, once per reading, and the reading dated, because answers drift between runs and over time; each answer recorded for whether the business was named, whether it was named first, and who else was named; the answers kept verbatim; a person checks each report before it is sent; the result a visibility score from 0 to 100, the share of the twelve questions that named the business, measured again on the same questions once the fixes are live, and every month with monitoring.

You can run a version of this yourself. Write five questions your buyers would type, save them, and once a month put each to the four assistants and record in a spreadsheet whether you were named, in what position, and which competitors were. Do not change the wording between months. The first run is the baseline, and everything after it is only meaningful against that. The number will move between runs on its own; what you are looking for is movement that persists across three runs and across months.

What a realistic first 90 days looks like

1. Weeks 1 to 2: measure. Run the baseline panel. Fetch your own pages the way a crawler does, with scripts off, and run the crawler access check. Write down what is missing in each of the four columns.
2. Weeks 3 to 4: identity and readability. One name, address and number everywhere, including the listings. Fix the bot wall, move text out of scripts and PDFs into HTML, and add LocalBusiness or the trade's type as JSON-LD with the fields filled.
3. Weeks 5 to 8: write the answers. The five to ten questions buyers ask on the phone, each as a heading with a plain answer under it, on the page where the buyer would look. Services by name, area, prices or ranges where you quote them, hours, credentials as facts. Reviews quoted on the page with attribution, a rating with its count.
4. Weeks 9 to 12: earn mentions. Bring the profile and the review sites into line with the site, ask recent customers for reviews where they already review, join the list your association publishes, and put one piece of your own knowledge somewhere your trade reads.
5. Week 13: re-measure with the same panel. Look first at the specific questions, the ones that name a service and a place, and expect the four assistants to differ. Two data points are a direction, not a trend; the third month tells you which.

Ninety days is enough to do the work and to see whether the assistants have picked it up. It is not enough to be sure. The assistants re-fetch pages on their own schedules, their answers vary between runs, and a competitor may be doing the same work. That is why the panel stays fixed and the measurement repeats.

The same plan, week by week and printable, is the 90-day GEO roadmap.

The honest limit: nobody can buy a place in the answer

There is no lever inside these systems that an agency can pull for you. gaash.ai is an independent agency with no arrangement with OpenAI, Anthropic, Google or Perplexity, and no agency has one. Google's page on AI features says AI Overviews need no special markup, file or optimization beyond ordinary SEO; the other assistants publish no path to a placement either. Anyone who guarantees you a place in an answer is guaranteeing something they do not control.

What can be done is what this page describes: make the facts about the business readable, attributable and corroborated, and measure whether the assistants pick them up. That is what our pricing is built around. The visibility check is free, and a person sends the result within 48 hours. The audit is $490 and is credited in full against the optimization work, which starts from $1,900 with a fixed scope. Monitoring is $290 a month and can be canceled in any month. None of it includes a promise of placement, because none could.