Why Doesn't ChatGPT Recommend My Brand?

It usually starts with a screenshot. Someone on the team asks ChatGPT which companies to consider in your category, the answer names three competitors, and the screenshot lands in a team channel with a question mark under it. By the afternoon there is a theory. We need more blog posts, or schema markup, or a Wikipedia page.

Maybe. But the screenshot proves less than it seems to, and the reasons ChatGPT leaves a brand out have very different fixes. A site ChatGPT can't read needs a developer. A brand nobody writes about needs PR. Treat one as the other and you can spend a quarter on the wrong problem.

This guide follows the order we use with customers, and with our own site: first make sure the problem is real, then find out which cause you have, then set up a way to tell whether the fix is working.

First, check that it's really happening

ChatGPT rarely gives the same answer twice. At the end of 2025, SparkToro and Gumshoe had about 600 volunteers put 12 recommendation prompts to ChatGPT, Claude and Google's AI, each one between 60 and 100 times. The odds that ChatGPT returned the same list of brands in any two of those answers were under 1 in 100. The same list in the same order came up roughly once in a thousand runs. One hospital in their data appeared in 69 of 71 ChatGPT answers and was named first in only 25 of them.

So a single answer, yours or a colleague's, is one draw from a wide spread of possible answers. Your brand can appear in four answers out of ten and still be missing from the one that got screenshotted. The reverse holds too: one flattering answer doesn't mean much.

Two more things distort a quick test. ChatGPT personalizes: OpenAI's help pages say product results take your context into account, including memory and custom instructions. A marketer who has spent months asking ChatGPT about their own industry isn't seeing what a stranger sees. And people phrase the same need in very different ways. In SparkToro's follow-up, 142 people asked for the same thing, headphones for a relative who travels a lot, and their prompts had an average semantic similarity of 0.081: they barely overlapped, and hardly anyone would have typed the prompt you tested.

A fairer test takes an hour. Write 20 to 30 prompts the way your customers would ask them: category questions, "best X for Y", comparisons, "alternatives to [competitor]". Run each one several times in a temporary chat, or logged out, and count how often you appear. That share is the number to watch, not your position in any one list. If you're in 40% of answers, you don't have a "ChatGPT doesn't recommend us" problem. You have a consistency problem, which is a different and smaller one.

Two ways ChatGPT knows about you

Everything below depends on one distinction. ChatGPT can answer from what the model learned during training, or it can search the web and answer from what it finds. Often it does some of both.

It searches less often than people assume. Semrush's clickstream panel, more than a billion lines of US browsing data from October 2024 to February 2026, found web search switched on for 34.5% of ChatGPT queries in February 2026, down from 46% at the end of 2024, with the monthly share swinging anywhere between 15% and 66%.

The rest of the time the answer comes from the model's memory, which reflects the web as it was when the model was trained and only changes when a new model ships. If your brand is new, recently renamed, or simply wasn't written about much before that point, the model may not know you at all, and no work on your website will change that before the next release. When ChatGPT does search, different things matter: whether your pages can be fetched, whether they come up for the searches ChatGPT runs, and what they say when they do.

A quick way to see which side you're on: ask the same question twice in a clean chat, once as is and once with "search the web" added. If you appear only with search, the live web knows you but the model doesn't yet. If you appear in neither, read on.

Seven usual causes

1. ChatGPT can't read your site

This is the cheapest cause to rule out, and the most annoying one to find.

OpenAI runs separate crawlers. GPTBot collects training data. OAI-SearchBot is the one that decides whether your pages can appear in ChatGPT's search answers, and OpenAI's documentation is blunt about it: a site that blocks it will not be shown in ChatGPT search answers, apart from plain navigational links. Blocking GPTBot doesn't take you out of search, but plenty of sites block every crawler with "AI" in its name, or sit behind a firewall or bot-protection setting that turns OpenAI away before robots.txt is even read. OpenAI says a robots.txt change takes about a day to reach its search systems.

The second problem is JavaScript. When Vercel and MERJ studied AI crawler traffic across Vercel's network in December 2024, none of the major AI crawlers, OpenAI's included, executed JavaScript. They read the HTML your server sends and nothing else. If your product descriptions, prices or articles only appear after a script runs in the browser, the crawler gets an empty page. That study is nearly two years old and crawlers can change, but we wouldn't bet a website on it.

We found a version of this on our own site this week. Lens's site audit flagged our blog index as unreadable. In a browser the page looked fine. The HTML our server sent to crawlers, though, held nothing but a list of post titles, about 700 characters of text wrapped in a JavaScript app, which is exactly what a script-rendered page looks like from the outside. The fix took an afternoon: crawlers now get each post's summary and date as well.

How to check: open your robots.txt and look for rules covering OAI-SearchBot and GPTBot, and for blanket Disallow lines. Ask whoever runs your hosting or CDN whether bot protection blocks AI crawlers. Then look at a key page without JavaScript ("view source" rather than the browser inspector, or curl with a crawler user agent) and check that the content is actually in the HTML.

2. You don't show up in the searches ChatGPT runs

When ChatGPT searches, it doesn't search for the prompt as typed. It rewrites it into a few narrower queries, often called query fan-out, and sends those to its search providers. OpenAI's help pages say it "typically rewrites your query into one or more targeted queries", but they don't name the providers or explain how results are ranked.

What happens after that has been studied closely. Ahrefs analyzed 1.4 million prompts on ChatGPT 5.2 (published April 2026) and found ChatGPT cites only about half of the URLs it retrieves. The pages it kept had titles that matched the fan-out queries more closely than the ones it dropped, and pages with readable URLs were cited more often (89.8%) than pages with auto-generated ones (81.1%). ChatGPT makes a quick call from the title, the URL and the snippet, much like a person scanning a page of results.

Which search index matters most is less settled than many guides suggest. In February 2025, Seer Interactive compared about 500 ChatGPT search citations with search results for the same questions and found that 87% matched Bing's top results, against 56% for Google. Ahrefs, counting differently, found only about 8% overlap between ChatGPT's in-text citations and the top ten of either Google or Bing. Different prompt sets, different ways of counting; we don't think either study closes the question. What looks consistent is that ordinary search strength helps. In a 2025 study using GPT-4o, Seer found a correlation of about 0.65 between ranking on Google's first page and being mentioned by ChatGPT. Being easy to find in conventional search, Bing included, which many teams have never looked at, is cheap insurance.

How to check: for your important prompts, guess what ChatGPT would actually search for. "Which project management tool is best for a small agency?" might become "project management software for agencies comparison 2026". Search those phrases on Google and Bing. If you're absent from both, ChatGPT's search step is unlikely to find you. If you've never set up Bing Webmaster Tools, do it; it takes ten minutes.

3. The web doesn't talk about you

For recommendation questions, ChatGPT leans heavily on pages where someone else compares the options. Ahrefs looked at what ChatGPT cited for 750 top-of-funnel prompts and found that "best X" list articles made up 43.8% of the pages. Wix's analysis of 75,000 AI answers points the same way: on commercial queries, list articles took about 41% of citations, close to double any other format.

This is the cause behind most of the cases we see. A company has a good product, a decent website and reasonable search rankings, and it's missing from every roundup, review site and comparison in its category. Whether the model is answering from memory or from search, it almost never sees the brand named next to its rivals, so it doesn't put it there either. We've written before about why mentions matter more than links and about which sources AI assistants actually read.

The broader evidence fits, with a caveat. In Ahrefs' study of 75,000 brands, how often a brand was mentioned across the web correlated much more strongly with appearing in Google's AI Overviews (0.664) than its backlinks did (0.218). That study is about Google, not ChatGPT, and correlation isn't causation. Still, it matches what we see: you get into AI answers by being talked about, not by being linked to.

How to check: search your category's "best of" lists, review platforms and comparison articles, in the language your customers use, and count how many name you. Then note which pages ChatGPT cites when it answers your prompts. In practice, the people who wrote those pages are deciding whether you get recommended.

4. ChatGPT isn't sure what you are

Open a clean chat and ask: "What is [your brand], and who is it for?" The answer tells you a lot.

If it mixes you up with another company of the same name, describes a product you dropped, or puts you in the wrong category, you are not one clear thing to the model. This happens to brands with common-word names. It happens to companies whose websites talk in abstractions ("solutions that empower growth") instead of saying what they sell and to whom. And it happens to brands that describe themselves one way on their website, another on LinkedIn and a third way in press coverage.

Recommendation prompts are about categories and needs: "a CRM for real estate agents", "a skincare brand for sensitive skin". If nothing written about you says you are a CRM for real estate agents in roughly those words, the model has no reason to connect you to that question, however good the product is.

The fix isn't glamorous. Describe yourself the same way everywhere, in the plain words customers use. Your homepage, about page, product pages, LinkedIn, directory listings and the boilerplate at the bottom of your press releases should agree on what you are and who it's for. If you also measure how AI describes you over time, brand perception is the place to watch.

5. The answer came from sources in another language

Outside the English-speaking world this one matters a great deal, and it's easy to miss.

In 2026, Peec AI analyzed more than 10 million prompts and around 20 million fan-out queries from its platform, and found that ChatGPT often runs its background searches in English even when the user writes in another language. Across non-English prompts, 43% of the fan-out queries were in English. For Turkish prompts, 94% of sessions included English searches, the highest rate of any language in their data. In Peec's examples, those English searches surfaced English-language sources that favored global brands over local ones. The data comes from one vendor's platform and the examples are individual queries, so we'd treat the exact numbers carefully, but the direction is hard to ignore.

For a strong local brand the consequence is uncomfortable. You can lead your market in your own language and still be invisible in the English sources ChatGPT reads to answer a question asked in that language. In Turkey, where ChatGPT accounts for 94.49% of AI-driven referrals, that's a large share of the market looking through one English-language window. A clear English page about what you do and where, and coverage in English-language comparison or industry sources, close part of the gap.

6. Your best pages are out of date

AI assistants lean towards newer pages. In Ahrefs' study of about 17 million citations (July 2025), pages cited by AI assistants were on average 25.7% newer than pages ranking in Google's organic results, and ChatGPT had the strongest preference of the assistants studied, citing URLs 393 to 458 days newer than Google's organic results.

Freshness isn't everything. The median page ChatGPT cited in Ahrefs' later study was still about 500 days old. But if your comparison page still says 2024, lists features you no longer offer and leaves out the competitor that launched last year, a newer page will tend to take the citation. Show a real date on pages and update them when something actually changes, not just to bump the date.

7. ChatGPT knows you, and didn't like what it read

Sometimes the brand is in the answer, but as a caution: "some users report slow support", "more expensive than the alternatives", or a comparison in which a competitor comes out ahead. Sometimes it's left out altogether because what the model found about you was negative.

Ask the questions a sceptical buyer would ask: "What are the downsides of [brand]?", "Is [brand] reliable?", "[Brand] or [competitor], which should I choose?" If the same complaints keep coming back, find where they start. Usually it's a handful of review pages, a forum thread or an old news story. What works, slowly, is fixing the problem underneath and getting newer, accurate material published; arguing with the model doesn't.

If you sell products online

Shopping answers work a little differently. OpenAI says product results in ChatGPT are chosen independently and are not ads, and that merchants are ranked on factors such as availability, price, quality and whether they are the maker or the primary seller of the item. Product data comes from third-party providers or straight from merchants. Shopify and Etsy catalogues are already connected, and other merchants can apply to send a direct product feed. Shopping is rolling out country by country, so check whether it's live where you sell. This is where clean product data pays off most: titles that say what the product is, correct prices and stock, real reviews.

What usually doesn't help

Buying ads. ChatGPT now sells ad placements in many countries, but the ad sits below the answer and doesn't change what the answer says. We looked at this in detail in our post on ChatGPT ads.

Writing for the model instead of the reader. The 2024 Princeton-led paper that named generative engine optimization tested a range of ways to rewrite a page. Keyword stuffing, the old search habit, gave little to no improvement. What helped most was citing sources, adding quotations and adding statistics, which raised visibility by 30 to 40% in their tests (a best case, not a typical result). The common thread is making the content more useful and easier to check, which is also what a human reader wants.

Schema markup, an llms.txt file or a new FAQ section are often worth adding too, but on their own they rarely explain why a brand is missing, which is why we'd rule out the seven causes above before reaching for them.

How to track whether it's working

The second half of the question matters as much as the diagnosis, because most fixes take weeks to show and some wait for a new model.

Track the share of answers that name you, not your position. Given how much answers vary, rank in a single list is noise, which is also why we're sceptical of tools that report an "AI ranking". Rand Fishkin's advice after the SparkToro study was to run each prompt 60 to 100 times. You don't need that many to see a trend, but you do need more than one.

Keep a fixed set of prompts. Choose the 20 to 50 questions closest to your revenue and leave them alone, so that a change in the number reflects a change in the answers and not in what you asked. Add new prompts next to them rather than swapping them in.

Look at each AI platform on its own. ChatGPT, Gemini, Claude, Perplexity and Google's AI Overviews read different sources, and a gain on one says little about the others.

Keep the answers, not only the score. When the number moves you'll want to know which competitor took your place and which source the answer relied on.

And expect a lag. A page fix can show up within days on prompts where ChatGPT searches. A change in how the web talks about you takes months. A change in what the model knows without searching waits for the next model.

This is the job Lens was built for. It asks your questions on each AI platform, keeps every answer, measures how often you're named next to the competitors you track, and turns what it finds into a list of actions, including the technical problems from the first cause, which it checks on your page once you mark them fixed. If you'd rather run the diagnosis by hand first, the steps in this guide are enough to start.

A short checklist

1. Run 20 to 30 real customer prompts several times each, in a clean chat, and count how often you appear.
2. Compare answers with and without web search.
3. Check robots.txt and bot protection for OAI-SearchBot, and load key pages without JavaScript.
4. Search the queries ChatGPT is likely to run, on Google and on Bing.
5. Count the "best of" lists, reviews and comparisons in your category that name you.
6. Ask ChatGPT what your brand is, and make your description consistent wherever it differs.
7. If your market isn't English-speaking, check what English-language sources say about you.
8. Update comparison and product pages that have gone stale.
9. Ask the sceptic's questions, and trace recurring complaints to their source.
10. Track your share of answers over time, per platform, on a fixed set of prompts.

Sources

• SparkToro & Gumshoe — "AIs are highly inconsistent when recommending brands or products", Rand Fishkin, January 2026 (data from November–December 2025)
• OpenAI — Crawler documentation (OAI-SearchBot, GPTBot, ChatGPT-User); Help Center, "ChatGPT search" and "Shopping with ChatGPT Search"; "Using shopping research in ChatGPT"
• Semrush — ChatGPT traffic analysis, 17 months of US clickstream data, October 2024–February 2026
• Vercel & MERJ — "The rise of the AI crawler", December 2024
• Ahrefs — "Why ChatGPT cites one page over another (study of 1.4M prompts)", April 2026; "Do AI assistants prefer to cite fresh content?", 17 million citations, July 2025; "An analysis of AI Overview brand visibility factors (75K brands studied)", 2025; ChatGPT and Google citation overlap analysis, 2025
• Seer Interactive — "87% of SearchGPT citations match Bing's top results", February 2025; "What drives brand mentions in AI answers?", 2025
• Wix Studio AI Search Lab — Content types most cited by LLMs, 75,000 AI answers
• Peec AI — "ChatGPT searches in English, even when you don't", 2026
• Aggarwal et al. — "GEO: Generative Engine Optimization", KDD 2024
• We Are Social & Meltwater — "Digital 2026 Global Overview"