← BlogExplainer

How ChatGPT chooses which businesses to recommend, explained

How ChatGPT chooses which businesses to recommend: what it reads on the web, what it remembers, the five things that shape an answer, and why answers vary.

ChatGPT chooses which businesses to recommend from two places: the pages it reads on the web while it answers, and what it learned in training. It names the businesses those sources describe clearly and consistently as a fit for the question.

There is no ranking you can look up, and no setting that adds you to an answer. What you can shape is the sources it reads. This guide explains both sources and the five things that shape who gets named. Then it covers why the same question gets different answers, and what all of it means for your own pages.

Where a ChatGPT answer comes from

From the web

For many questions, ChatGPT searches the web while it answers. It reads the pages it can fetch and writes its answer from them, often with links to the pages it used.

OpenAI's page on its crawlers, read on September 29, 2026, names the visitors involved. OAI-SearchBot is used to surface websites in ChatGPT's search features. ChatGPT-User may visit a page when a person asks ChatGPT a question. The same page says a site that blocks OAI-SearchBot is not shown in ChatGPT search answers.

This is the path your pages can affect directly. If a page answers the question and ChatGPT can fetch it, it can be read the next time someone asks.

From memory

ChatGPT also answers from what it learned in training, before your question was asked. That memory reflects how the web described businesses when the model was trained. OpenAI's same page says GPTBot collects content that may be used to train its models.

You cannot edit that memory. What you can shape is the web the next model learns from: your own pages, and what other sites say about you.

Which path an answer took

You often cannot tell, and many answers mix both. Links to sources are one sign that ChatGPT searched. An answer with no links can still draw on pages the model read in training. For your work, the difference matters less than it seems, because both paths reward the same kind of page.

What shapes how ChatGPT chooses which businesses to recommend

Five things decide whether your business is one of the names. None of them is a hidden setting. Each is something you can check and change.

1. A page answers the question

ChatGPT writes answers from sentences it can lift. A page that answers the question in its first lines, in words close to how customers ask, gives it that sentence. A page that talks around the question gives it nothing to use. A service page that only says "we do it all" does little for "who can fix a burst pipe on a Sunday?"

2. ChatGPT can read the page

A page ChatGPT cannot fetch does not exist for that answer. Three things stop it: settings that turn its visitors away, text that only appears after scripts run, and answers locked inside images. Rankli's free AI check tests your site for these and lists the fixes, with no account needed. To see which businesses it names for your own market, the free ChatGPT visibility checker asks it three buyer questions and shows the answers.

3. The facts agree across sources

When your site, your listings and other people's pages agree on what you do, where and for how much, there is one story to repeat. When they disagree, the answer can repeat the wrong version. A price that changed on your site but not on a directory is one way this happens.

4. Other sources name you for the same thing

ChatGPT reads more than your own site. Review sites, directories, comparison pages, forums and local news all describe businesses. A business that several independent sources name for the same need gives ChatGPT a consistent picture. A claim that only your own site makes gives it one source to draw on.

Which sources count depends on what you sell. A local business is described by directories, review sites and local news. A software product is described by comparison pages, review sites and the forums where buyers ask for tools. A shop is described by product reviews and buying guides. Find the sources that describe your rivals, and make sure they describe you too, with the same facts your own site states.

5. The question itself

The wording decides the shortlist. "Best accountant" and "accountant for a small online shop in Portland" get different answers, because the second names a need, a type of customer and a place. Your pages should say plainly which needs, customers and places you serve. Then ChatGPT can match you to the questions where you fit.

What ChatGPT needs to know about you

An answer that recommends a business has to say why. So ChatGPT needs a few plain facts it can state with confidence. Check that your own site says each of them in words, on a page it can read:

  • What you sell, named the way a buyer would name it, not in your internal terms.
  • Who it is for. The type of customer, their size or situation, and the need you solve.
  • Where you work, for any business that serves a place: the city, the neighborhoods, the area you cover.
  • What it costs, or at least a starting price and what it includes.
  • How you differ from the obvious alternatives, stated as facts a buyer can weigh.
  • How to buy or book. A clear next step, linked in plain HTML.

If a fact is missing, the answer has nothing to say about it, and a rival whose pages state it can be named instead. Each missing fact is one sentence to add.

Why the same question gets different answers

Ask ChatGPT the same question twice and the names can change. There are three reasons:

  • Each answer is written when you ask. It is generated for that asking, so the names can differ from the last one.
  • The web changes. New pages appear and old ones change, so a search today can find different pages from last week.
  • Small wording changes matter. "Best" and "cheapest" are different questions, and so are "near me" and a named city.

So one answer tells you little. Ask each question several times and count how often you are named. That count, tracked over time, is the real picture.

What does not help

Some moves feel like they should work and add nothing.

  • Repeating a phrase. Repetition adds no facts. An answer needs something to state: what you do, where, how much and for whom.
  • Claims only you make. "The best plumber in Austin" on your own site is one source. Rivals that many sources describe give the answer more to draw on.
  • Pages it cannot read. Pages built from images, or with text that loads late, arrive blank to a visitor that reads only the page as sent.
  • Publishing and never checking. Without asking the question again, you cannot tell whether a page changed anything.

How to see what ChatGPT chooses today

You can see it by asking. Write down the questions your customers ask before they choose. Put each one to ChatGPT in a fresh chat, several times, and record who it names and which pages it links. Does ChatGPT recommend your business? gives you a step-by-step method and a starter list of questions.

Rankli does this every week. It asks ChatGPT, Claude and Gemini the questions your customers ask, with Perplexity added on Scale and Agency. It shows the questions where AI recommends someone else, who it picks instead, and which pages each answer cites. The AI visibility page lists what each weekly round records.

What this means for your pages

Put the five factors together and the work is clear:

  1. Publish one page for each question you are losing, with the answer in its first lines.
  2. Make sure ChatGPT's visitors can fetch it, and that the words are in the page itself.
  3. Keep your name, services, places and prices the same on every page and listing.
  4. Be described by others: correct listings, comparison pages that include you, and reviews on the sites your customers use.
  5. Ask the questions again after each page goes live, and compare.

Rankli runs that loop for you. It writes an article aimed at each question you are losing, publishes it to your site, and asks again 7 and 14 days after it goes live. How to get recommended by ChatGPT walks through the same steps by hand.

Questions

Does ChatGPT rank businesses like a search engine does?

No. A search engine lists pages in order, while ChatGPT writes an answer that names a few businesses from the pages it reads and what it learned in training. Rankli measures that answer directly, by asking ChatGPT your customers' questions every week.

Why does ChatGPT recommend my competitor and not me?

The pages it read may describe your competitor more clearly for that question, or it may not be able to read yours. Rankli shows the questions you are losing, who wins each one, and the fixes that stop AI reading your site.

Does ChatGPT read reviews and directories?

When it searches the web, it can read any page it is allowed to fetch, including review sites and directories. Rankli shows which pages each answer cites, so you can see which sources shape it.

Can ChatGPT remember my business?

It can, if your business was described on the web when the model was trained. You cannot edit that memory, so Rankli works on the pages ChatGPT can read today, and asks again to see what changed.

Why do I get a different answer each time I ask?

ChatGPT writes each answer fresh, and the pages it finds can change. That is why Rankli asks the same questions every week, and three times on each assistant when it re-checks an article.

How this post was made

The Rankli team wrote this post, and the product scored it. It went through the same 15 rules a customer’s draft goes through: the word-count range, the heading structure, the keyword placement, the fabricated-statistic detector and the rest. Every number below came back from that run on the words above it.

100SEO score15 of 15 rules passed
Written by
Rankli editorial
Target keyword
chatgpt chooses which businesses to recommend
Words
1,660
Keyword density
1.08%
Reading grade
6.1
Internal links
5

Every rule, as it scored

hard blocks a draftadvisory lowers the score

  • Passed: word-counthard1660 words; the target range is 1500–3000
  • Passed: heading-hierarchyhardone H1, no skipped levels
  • Passed: keyword-in-titlehardtitle contains "chatgpt chooses which businesses to recommend"
  • Passed: keyword-in-openinghardthe first 100 words must mention "chatgpt chooses which businesses to recommend": searchers and crawlers both decide there
  • Passed: keyword-in-h2hardat least one H2 must contain "chatgpt chooses which businesses to recommend"
  • Passed: keyword-densityharddensity 1.08% of words; above 2.5% reads as stuffing: vary the phrasing
  • Passed: banned-phraseshardno banned phrases
  • Passed: no-fabricated-specificshardno unsourced statistics, citations or quotes detected
  • Passed: internal-linkshard5 internal link(s); 2–5 contextual links to the customer's own pages
  • Passed: subject-coveragehardmentions "answer", "engine", "rankli" outside the keyword
  • Passed: section-substancehardevery section carries at least 25 words
  • Passed: title-deliversadvisorythe body delivers what the title promises
  • Passed: meta-description-lengthadvisory155 characters; aim for 150–160 so it fills the snippet without truncation
  • Passed: title-lengthadvisory60 characters; titles over 60 are truncated in results
  • Passed: readabilityadvisoryreads at grade level 6.1; above 12, sentences are likely too long for a general audience

The rules are the ones in src/content/seo.ts, unchanged for us: hard rules block a draft from ever reaching a customer’s site, advisory rules only lower the score. Scored and committed as a markdown file, so the commit is the receipt. Run the same fifteen checks on a draft of your own in the SEO article checker, or see what it costs.

Rankli writes posts like this for your site.

Same fifteen checks. 7 days free, then your plan. Cancel before day 8 and nothing is charged.

Start free trial

Keep reading

All posts →