Getting recommended by ChatGPT: the six steps a business controls

Why an answer may have nothing to say about you

An AI assistant can only say about a business what it has read somewhere. When ChatGPT answers with a search, it fetches pages the way a crawler does, usually without running scripts, and what it can say about a business comes from what those pages state: who it is, where it is, what it does and what others say about it. A business whose pages leave those things out gives an answer nothing of its own to repeat.

Most small business sites are in that state. In our audit of 9,483 small business websites (40 trades, 90 metros, September 2026), 61.5% had no question-and-answer content of any kind. 43.1% were inconsistent about who and where they are across their own homepage, contact page and about page. 64.7% showed no review signal a crawler could read. Fewer than one in five (17.8%) had no schema.org markup, and on most of the rest the markup described the website, not the business.

Those are the four things, and they are the whole guide: readable, consistent, answer-shaped, reviewed. The median audited homepage ran to 880 words, so the problem is rarely that a site says too little. It is that what it says cannot be attributed, quoted or verified by a machine reading it.

Step 1: measure before you touch anything

A fix you cannot compare against a starting point is a guess. Before changing a line, write down what the assistants say today, in a form you can repeat in a month. This is how we measure a client, and one person can do a smaller version in two or three evenings.

1. Write ten to twelve questions the way your buyers type them: the problem question ("who fixes a furnace in Ottawa at night"), the comparison question ("most trusted family dentist near downtown Tucson"), the constraint question ("emergency plumber that takes card payment") and one that names you.
2. Put each to ChatGPT, Claude, Gemini and Perplexity, three times each. Answers vary between runs; the average is the part you can rely on.
3. Record four things per answer: named, recommended, first, and who else was named. Keep the answer verbatim so any verdict can be checked.
4. Date the sheet and leave the questions alone. That is your baseline, and every later reading compares against it.

If you would rather a person did this, the free visibility check takes a website and an email and sends a written result within 48 hours: whether an AI assistant, Claude with live web search, names you when a buyer asks, and who it names instead.

Step 2: make the site readable

Nothing else counts if the crawler cannot reach the words. OpenAI's crawler documentation describes three AI crawlers with different jobs: OAI-SearchBot indexes pages for ChatGPT search, ChatGPT-User fetches a page when a user asks about it, and GPTBot collects training data. Disallowing OAI-SearchBot removes a site from ChatGPT search answers; disallowing GPTBot does not. A robots.txt line written in 2023 to keep GPTBot out has no bearing on whether ChatGPT can cite you today.

1. Open /robots.txt and check that the answer-side tokens are allowed: OAI-SearchBot and ChatGPT-User for ChatGPT, Claude-SearchBot and Claude-User, PerplexityBot, Googlebot. The crawler access check reads the file for all nineteen tokens the way RFC 9309 says a crawler does, and our article on robots.txt for AI crawlers has a file you can copy. 3.2% of the audited sites disallow AI crawlers on purpose.
2. Fetch your homepage as a crawler would: curl -A "OAI-SearchBot" -I https://www.example.com/. A 200 is a page; a 403, a 503 or an HTML challenge is a bot wall, and 7.2% of the audited sites served our crawler one. The setting that raises it is in your CDN's or security plugin's bot management, where the answer-side tokens can usually be allowed by name.
3. View the source of a service page with scripts off. If the words you sell with are not there, move them there. Our crawler executed no JavaScript, none of the assistants' crawler documentation promises a rendering step, and Google's JavaScript documentation describes rendering as a separate, deferred step. 3.0% of the audited sites render their text only in scripts, and 2.1% keep it in a PDF.

Step 3: one identity

State one name, one address, one phone number and one set of hours, the same everywhere they appear. If your pages carry two phone numbers, an answer that gives one may give the wrong one. 36.9% of the audited sites publish more than one, and 34.1% have no postal address a machine can parse at all.

1. Settle the four facts and write them identically in the footer, on the contact page, on the about page and in your Google Business Profile, whose guidelines ask for the real name, the real address, the right categories and the real hours. The consistency audit is a worksheet for the check.
2. Add LocalBusiness schema, or the specific subtype for your trade (Dentist, Plumber, LegalService, AutoRepair), as JSON-LD on the homepage, with name, address, telephone, openingHoursSpecification, areaServed and url filled in. Google's local business page lists the properties it reads and schema.org lists the subtypes. 13.7% of the audited sites had business markup with those fields missing.
3. Make the markup match the visible page word for word. Google's guidance is to mark up only what a visitor can see, and JSON-LD is the format it recommends.

Step 4: write the answers

This is the step that gives ChatGPT something to quote, and it is the one most sites skip. 61.5% of the audited sites had no question-and-answer content; 50.7% had a heading phrased as a question with no answer under it a machine could isolate. Assistants lift answers from pages that already contain answers.

The mechanism has been measured outside this site. Aggarwal et al. (KDD 2024) tested changes to source pages and found that adding statistics, quotations and citations raised a source's visibility in generated answers by up to about 40% on their benchmark. Specific, attributable passages get lifted; slogans do not.

1. For each service, write the five questions buyers ask on the phone as headings, with an answer of 40 to 80 words directly under each: answer first, detail after.
2. Put the facts the buyer wanted inside the answer: the price or range, the area, the hours, how to book.
3. Give every service its own answer page. The answer-page template is the whole structure as a filled example.
4. Do not write a separate version for AI. Google's page on AI features asks for no special file, markup or optimization, only the SEO basics; write the page for the buyer who asked.

Step 5: reviews and mentions

64.7% of the audited sites show no review signal a crawler can read, and 8.9% carry rating markup. Most of these businesses have reviews. The site does not say so, and a fetch of the site returns the site, not the platform the reviews sit on.

1. Quote two or three real reviews on the page as text, with the reviewer's first name, the date and the platform, linked, and put each on the service page it concerns.
2. State the rating and the count in the text, with the platform named, and keep them out of the markup. Google's review snippet rules say not to aggregate reviews or ratings from other websites, and a business's reviews of itself on its own site earn no review stars.
3. List where your trade is discussed in your metro: the association's member directory, the two or three review sites buyers in your trade open, the neighborhood forum where people ask for a name. Get your listing on each correct and identical to the site.

The mentions themselves you cannot write. A brand mention is by definition text you did not produce: a customer naming you in a thread, a supplier listing you, a local paper quoting you. What you control is whether the name they use matches the one on your site, and whether a machine that follows the mention finds a page confirming it. In our probe, of every name the assistant gave, 99.4% were actual businesses and 0.6% were directories. The answer names businesses; the directories are one more place where the facts have to agree.

Step 6: re-measure monthly with the same panel

Run the same ten to twelve questions against the same four assistants, three times each, and fill in the same four columns. Compare question by question: an assistant that now names you for the comparison question but not for the problem question is telling you which page to work on next. Change the panel and the months stop comparing.

This is the part that continues. A visibility score is only useful as a series, and the series is only honest if the questions behind it never move. Our monitoring does this for $290 a month; a spreadsheet and a calendar reminder do it for nothing.

The six steps, who does them and how long they take

The 90-day shape

Put together, the steps fit in a quarter. Weeks 1 and 2 are the baseline: the questions written and run, nothing changed. Weeks 3 to 6 are the foundation: one identity everywhere, the LocalBusiness markup, robots.txt and the bot wall checked, the words in the HTML. Weeks 7 to 12 are the answers and the mentions: the five questions per service, the reviews on the page, the listings corrected. Day 90 is the re-measure with the same panel. The 90-day roadmap has every week on one page.

The honest limits

No operator sells placement in an answer, and no agency has a lever inside these systems. gaash.ai is an independent agency with no arrangement with OpenAI, Anthropic, Google or Perplexity. What this guide does is remove the reasons ChatGPT could not name you and give it something to quote once it can. What it cannot do is promise that it will.

One number is worth reading carefully. For 9,240 of the audited businesses we put one buyer question to Claude with web retrieval on and checked whether the business was named: 14.9% were, and 17.7% among the 5,928 answers that gave a differentiated list rather than a hedge. That is a Claude number, from one question per business on one day. It is not a ChatGPT rate, and we do not know what the same probe on ChatGPT would give. It is here because the businesses that were named were, on inspection, the ones whose sites had the answer-shaped content and the consistent identity described above.

So the claim is narrow. Measured this way, the businesses that were named tended to have the consistent identity and the answer-shaped content, and the six steps give you both, with the access and the reviews that make them count. Whether an assistant names you on a given day is its decision. The only honest way to know is to ask it, the same way, every month.