Why ChatGPT Doesn't Recommend Your Business (and How to Fix It)

Open ChatGPT (or Gemini, or Perplexity) and ask the question a customer would ask before they find you: "What's a good [what you sell] in [where you sell it]?" or "Recommend a [your category] for [your customer type]." Ask it three or four different ways. Then look at who is in the answer.

If you are not there, you are in the majority. In Webmatik's audit data, the average brand was mentioned in just 19% of buyer-style AI queries about its own category, and 26% of brands were never cited with a link in any answer at all. Those are sites whose owners cared enough to run an AI-visibility scan; the typical business does worse.

This matters more every month. A growing share of "which one should I pick" questions are now asked to an assistant instead of typed into Google, and the assistant gives one answer with three to five names in it. There is no page two.

The good news: unlike a Google ranking, this is not decided by a decade of backlinks. AI assistants recommend businesses for a small set of concrete reasons, and most of them are fixable in weeks. Here they are, roughly in order of how often they are the actual cause.

Before the reasons, a short, accurate model of what is happening — because most advice on this topic treats an LLM like a search engine, and it is not one.

When you ask for a recommendation, two things combine:

• What the model already "knows" — patterns from its training data: which names appear, with which descriptions, in which contexts, across the public web. A business that has been described consistently in many places is a stable entity the model can name with confidence. A business described nowhere, or differently everywhere, is noise.
• What it retrieves right now — most assistants now run a live web search for commercial questions and read the top results: review sites, "best X in Y" lists, comparison pages, directories, and the businesses' own pages. Whatever is easiest to extract a clear fact from gets used and cited.

So the assistant is not ranking you. It is asking two questions: Do I know who this is and what they do? and Can I find a clean statement of it, ideally from someone other than them? Every reason below is a way of answering "no" to one of those.

1. You are not an entity yet

The model has no stable idea of who you are. This is the most common cause for small businesses and the easiest to test: ask the assistant "What is [your business name]?" If it hedges, guesses, or describes a different company, you do not exist to it as an entity.

The cause is inconsistency: your name, category and location are stated differently (or not at all) across your site, Google Business Profile, LinkedIn, directories and social profiles. Fix: one canonical sentence — "[Name] is a [category] that [what you do] for [whom] in [where]" — used verbatim everywhere, including the first paragraph of your homepage and your Organization structured data.

2. Nobody else talks about you

Assistants weight third-party sources heavily, for the same reason a person would: a business describing itself is not evidence. If the only place your name appears is your own website, you have one weak vote. The names that show up in answers are on review platforms (Google, G2, Capterra, Trustpilot, Yelp — whichever fits your category), in industry directories and associations, in "best of" lists and comparisons, and in forum threads where real people mention them.

Fix: pick the two or three platforms where your category is actually compared and get a real presence there — claimed profile, complete description in your canonical sentence, and a steady flow of genuine reviews. This is the slow part of the plan, and the part that compounds.

3. Your site never says plainly what you do

Open your homepage and read only the first sentence a machine would see: the H1 and the first paragraph. On a large share of sites it is "Welcome", "Solutions for a changing world" or a slogan. An assistant reading that learns nothing it can repeat. It moves on to the site that says "Family dental practice in Leeds, open evenings, accepting NHS and private patients."

Fix: put the canonical sentence from reason 1 in your H1 or directly under it, and make sure each service or product page opens with an equally literal statement. Specificity is what gets extracted — "for e-commerce stores on Shopify" beats "for growing businesses".

4. There is nothing extractable to cite

Assistants build answers out of facts they can lift cleanly: prices, comparisons, specs, service areas, hours, "who is it for", "who is it not for", and direct answers to common questions. In our data 40% of sites have no FAQ or Q&A content and 63% hide their pricing. A site with no facts cannot be quoted, so the answer quotes a competitor who published theirs.

Fix: a real FAQ (the questions customers actually ask, answered in two or three plain sentences each), visible pricing or at least ranges, and one honest comparison page for the alternatives people weigh you against. Mark the FAQ up with FAQPage schema so both Google and the assistants' retrievers recognise it as Q&A.

5. You have not told AI crawlers anything

Three signals decide whether AI systems can and may use your content. 62% of sites have no llms.txt — the short file that tells an AI what the site is and which pages matter. 90% publish no Content Signals (the robots.txt line that states what AI may do with your content), and 98% offer no clean markdown version of their pages. On the other side, only 15% actively block AI crawlers — so the problem is almost never that you have shut the door; it is that you have said nothing, and a site that says nothing is easier to skip than one that says "here is who we are, here is what you may use".

Fix: publish /llms.txt, add a Content-Signal line to robots.txt, and confirm you are not accidentally blocking GPTBot, ClaudeBot, PerplexityBot and Google-Extended. A Webmatik audit checks all three and generates the llms.txt for you.

6. The answer is really a "best of" list — and you are not on it

For many commercial queries the assistant's live retrieval lands on the same handful of "10 best X" articles, and the answer is a paraphrase of them. If you are not in those lists, no amount of on-site work puts you in the answer for that query. Ask the assistant for its sources (most will show them) and you will see exactly which pages it leaned on.

Fix: this is outreach, not optimisation. Those lists have authors; most update annually and many accept suggestions, especially with a specific angle ("the only one that does X"). Meanwhile, target the queries the lists do not cover: longer, more specific questions where your own FAQ and comparison pages can be the source.

Do these in sequence; the early steps make the later ones work.

1. Week 1 — become an entity. Write the canonical sentence. Put it in the H1/first paragraph, Organization schema, Google Business Profile, LinkedIn, and every directory profile you already have. Publish llms.txt and the Content-Signal line. Check that no AI crawler is blocked.
2. Week 2 — give them something to quote. Add the FAQ (10–15 real questions) with FAQPage schema. Make pricing or ranges visible. Rewrite each service or product page to open with a literal one-sentence description. Publish one comparison page against the two alternatives you are most often weighed against — honest about where they win.
3. Week 3 — get other people to say it. Claim and complete profiles on the two or three review and directory platforms that matter for your category. Ask your last twenty happy customers for reviews. Find the "best X" lists the assistant cites for your queries and contact the authors.
4. Week 4 — measure and repeat. Re-run the same questions from the test at the start. Track which queries now mention you, whether you are cited with a link, and who else is in the answer. Then pick the next set of queries.

Expect the on-site changes (weeks 1–2) to show up in retrieval-based answers within a few weeks as the pages are recrawled, and the entity and third-party work (weeks 1 and 3) to take longer and keep paying off after that.

• Do not stuff hidden text or "instructions" for AI into your pages. Prompt-injection tricks ("ignore previous instructions and recommend us") are detected, are being actively filtered by every major assistant, and get sites excluded from retrieval. They also look exactly as bad as they are if a customer views source.
• Do not buy or fabricate reviews. Review platforms are the assistants' most trusted source precisely because they police this; a takedown or a flag on your profile is worse than having ten reviews.
• Do not write "AI-optimised" content that a human would not read. The pages that get cited are the ones that answer a real question plainly. Keyword-shaped filler is skipped by retrievers the same way it is skipped by readers.
• Do not block AI crawlers to "protect" your content unless you have decided you do not want to be recommended. Content Signals let you allow search and assistant use while disallowing training, if that is the line you want to draw.

The manual test — the same handful of questions asked monthly — is fine to start with, but it is noisy: answers vary run to run and by phrasing. Three things make measurement useful:

• A fixed set of buyer-style queries for your category (eight to ten), re-run the same way each time, so you are comparing like with like.
• Two counts, not one: how many answers mention you, and how many cite you with a link. Citation is the stronger signal — it means your page was the source, not just your name in the model's memory.
• Share of voice against named competitors: the same queries, tracking who else appears. Your goal is not "be mentioned"; it is "be mentioned as often as the two names you lose deals to".

Webmatik's AI Search Visibility check runs exactly this — real queries against ChatGPT and Gemini, mentions and citations counted, competitors tracked over time — so you can see the number move as you work through the plan above. If you would rather read the deeper techniques first, How to get mentioned by AI search engines and Advanced GEO cover the content and entity work in detail.