# Why AI doesn't recommend your product

Five causes, in order: an unreadable site, no place in the AI's live search, no third-party mentions, a vague homepage, weak reviews. A test for each.

By Dasha. Updated October 31 2026. https://heardofyou.io/guides/why-ai-doesnt-recommend-you

If ChatGPT, Claude, Gemini or Perplexity never name your product, it is almost always one of five things. They happen in this order:

1. The assistant cannot read your site.
2. You are not in the web results it pulls in before answering.
3. Nobody else has written about you where it reads.
4. Your homepage does not say what you are and who you are for.
5. A competitor beats you on reviews in a close call.

Below is each cause, the evidence, and a test you can run today. This is the deeper companion to [why ChatGPT doesn't recommend me](/why-chatgpt-doesnt-recommend-me).

## Cause 1: the assistant cannot read your site

First, the assistant's crawler has to fetch your pages and get readable text back. Three things quietly stop that.

- **A robots.txt rule that blocks AI crawlers.** OpenAI's search crawler is OAI-SearchBot, its training crawler GPTBot. Anthropic's are Claude-SearchBot and ClaudeBot. Perplexity's is PerplexityBot.
- **A firewall or bot-protection page** that serves a challenge instead of your content.
- **A site that only renders its words with JavaScript.** The first fetch returns an empty shell.

Brands that block Claude's crawler get almost no Claude citations (Muck Rack).

**Test today:** open your site's `/robots.txt`. Look for `Disallow` lines under any bot name above, or under `User-agent: *`. Then fetch your homepage as a bot would, with no JavaScript. Check that your product name and description appear in the raw text. The commands are on our [AI crawler check](/tools/ai-crawler-check) page.

## Cause 2: you are not in the assistant's live search results

This is the big one. When the assistant searches the web, it works from the pages that come back. ChatGPT searches in about a third of prompts, and adds the year to 76% of software searches.

In Tannenbaum's 2026 study, a brand missing from those results got mentioned by GPT 2.8% of the time. When the brand's own site was cited and branded searches happened too, 91.4%.

**2.8% : how often GPT named a brand that was absent from its live search results, against 91.4% when its site was cited : Tannenbaum, 2026 (pre-print)**
Position inside those results matters too. 58% of pages at position 1 got cited, and 14% at position 10 (AirOps and Kevin Indig, 815,484 query-page pairs).

**Test today:** write down the three or four searches the assistant would run for your buyer question. They look like "best [category] 2026", "[category] reviews", "[category] comparison" and "[your name] reviews". Search each one signed out, in a normal search engine. If no page about you shows up on page one, you found the cause.

You can also ask Perplexity your buyer question and expand its sources. It shows exactly what it retrieved.

## Cause 3: nobody else has written about you

Assistants prefer pages the brand does not control. For US software questions, 72.7% of cited sources were earned media (Chen et al.). Across 207 million citations, over 84% were third-party and 0.3% were paid (Muck Rack).

A product that lives only on its own site and a few directories is invisible in that mix. Directories are rarely cited (GetIntel, 9,978 answers).

**Test today:** search your product name with your own domain excluded, as in `"Ledgerly" -site:ledgerly.example`. Count real pages: reviews, "best X" lists, Reddit or forum threads, YouTube videos, articles. Fewer than five is normal for a solo product. On its own, it explains a near-zero mention rate.

Then run your buyer question through one assistant and note every source it cites. Those are the pages you are not on.

## Cause 4: your homepage does not say what you are

Even when the assistant finds you, it needs a clear sentence to quote. What the product is, who it is for, what it costs.

44.2% of citations come from the first third of a page. Definite wording gets cited more often (AirOps and Kevin Indig). Buyers need it too: 58.8% of AI referrals land on the homepage (Similarweb). The person who arrives wants to confirm what the assistant said, in one glance.

**Test today:** open your homepage and read only what shows before you scroll. Finish this sentence from it alone: "[Product] is a [category] for [audience] that [main job]; it costs [price]." If you cannot, neither can the assistant.

Then compare the assistant's description of you with the truth. A wrong price or audience means the pages it read are out of date.

## Cause 5: you lose the close calls on reviews

When the assistant weighs you against a bigger name, familiarity wins ties. With identical specs, the well-known brand was recommended 100% of the time (Chu and Hou, 2026).

The same study found the fix: a competitor rating edge of under 0.1 stars broke that lead. In Columbia's agent study, a 0.1 rating increase moved a product from a 10% pick rate to 15 to 20%. A reviewer "pick" badge took it to 20 to 43%.

**Test today:** find the review site the assistant cites most for your category. Often that is G2, Capterra, Product Hunt, Trustpilot or an app store. Put your rating and review count next to the two competitors it names most. No profile, or three reviews against their three hundred, is the cause.

Honest reviews from real customers, asked for plainly, are the lever a small product can actually pull.

## Two things that are not your fault

**Order and position.** Assistants react to the order candidates appear in what they read (Bito et al.). Agents favour items in the top row (Columbia). If every list puts the incumbent first, you inherit that bias. You cannot fix it directly. Being on more lists, higher up, is the only route.

**Noise.** The same prompt gives the same brand list less than 1 time in 100. Named last month and not this month? Run it three times before deciding anything changed.

## Limits

The 2.8% and 91.4% figures come from one pre-print, on one model family. The exact numbers will not hold everywhere, but the direction has held across studies. Muck Rack and GetIntel sell PR and visibility tools, so treat their figures as large but vendor-reported.

Fixing all five causes raises your odds. Nobody can promise what an assistant will say.

## What to do next

Run the tests in order and stop at the first one that fails. That is your cause.

It is usually cause 2 or 3. Then the work is getting named on the pages the assistant already cites for your category. That is what our project does. The [free check](/check) is where it starts: it tells you which pages those are.

## Questions

**ChatGPT named me once last month and not this month. Did I lose something?**

Probably not. Answers at the edge of the list change almost every run. Measure with repeated runs before you conclude anything.

**Is it because I have no llms.txt or schema markup?**

Almost certainly not. No evidence shows either one makes an assistant recommend you. The causes on this page are about whether it can find and read pages about you at all.

**I am on 40 startup directories. Why doesn't that help?**

ChatGPT and Gemini rarely cite directories. In a study of 9,978 answers they were a minor source. Reviews, best-of lists and community threads get cited instead.

**Which cause should I fix first?**

Work down the list in order. If the assistant cannot read your site, every later fix is pointless. Absence from live search explains most missing mentions.


## Sources

- Tannenbaum, A fitted stage model of brand mention in AI search, 2026 (pre-print) (https://arxiv.org/abs/2609.23162)
- AirOps and Kevin Indig, 815,484 query-page pairs (https://www.getpassionfruit.com/research/is-query-fan-out-a-marketing-buzzword-or-something-relevant)
- Chen et al., University of Toronto, earned media in AI search, 2025 (https://arxiv.org/html/2509.08919)
- Muck Rack, Generative Pulse, 207 million citations, October 2026 (https://www.globenewswire.com/news-release/2026/10/28/3390134/0/en/generative-pulse-earned-media-still-drives-more-than-84-of-ai-citations.html)
- Chu and Hou, brand bias and rating edges, 2026 (pre-print) (https://arxiv.org/abs/2606.17443)
- Columbia, What Is Your AI Agent Buying? (https://ar5iv.labs.arxiv.org/html/2508.02630)
- Bito et al., position bias in LLM recommendations (https://arxiv.org/abs/2508.02020)
- StatWharf, ChatGPT query fan-out research (https://statwharf.com/research/chatgpt-query-fan-out/)
- Similarweb, 2026 generative AI landscape (58.8% of AI referrals land on homepages) (https://www.similarweb.com/corp/reports/2026-generative-ai-landscape/)
- OpenAI, bots and crawlers documentation (https://platform.openai.com/docs/bots)
- Making a website LLM-friendly (research summary used for the crawl-access section) (https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- GetIntel, 9,978 AI answers: directories are rarely cited (https://getintel.ai/blog/best-saas-directories-2026/)