Your competitor gets named in AI answers and you don’t – what they have that you don’t

Usually it is not their website outranking yours. Across 121 AI answers to 24 commercial and comparison-intent prompts polled on ChatGPT and Gemini between 1 and 11 September 2026, the average answer cited 8.5 sources drawn from 541 distinct domains — and in the answers that did cite us, our own site supplied only 18.1% of those sources.

At a glance

• Confirm before you act. One answer is not evidence. Of the 25 prompt-and-engine pairs polled two or more times between 3 July and 11 September 2026 that cited us at least once, 23 also failed to cite us on another poll of the same prompt on the same engine.
• Audit the sources, not the answer. The question is not “who did it name” but “which pages did it read”.
• Own-site content is the minority input. In our own measured window, 82% of the sources behind the answers that named us were on domains we do not control.
• Engines disagree. ChatGPT drew 24.1% of its citations from non-vendor sources in that window; Gemini drew 7.6%.

“Our competitor is cited by AI and we are not” — confirm it before you spend a dollar

Most of these investigations start with a single screenshot, and a single screenshot is close to worthless. On our own citation monitor — a registry of 132 active prompts, 71 of which were polled on ChatGPT or Gemini in this window — we looked at every prompt-and-engine pair polled two or more times between 3 July and 11 September 2026: 66 pairs, 1,309 polls. Twenty-five of those pairs cited LeadsNow at least once. Twenty-three of those 25 — 92% — also produced at least one answer that did not cite us, same prompt, same engine, days apart.

On the prompts where a brand is genuinely present, it still disappears from individual answers routinely. If you asked once and saw a rival named, you measured a coin flip, not a competitive position.

The five-poll rule: how many times to ask before you believe the gap

Among those 25 prompt-and-engine pairs, the median per-poll citation rate was 57%. Treat that as the behaviour of a brand that is in the engine’s consideration set, and the arithmetic for a single check is unforgiving: you miss it 43% of the time. Repeat the poll and the miss probability compounds — 0.43 × 0.43 for two, and so on. This assumes polls are independent, which is an approximation: answers drift with the index rather than being redrawn cleanly each time, so treat the table as a floor on how much asking you need to do, not a ceiling.

The five-poll rule: poll the same prompt on the same engine five times across at least three days before you accept that a competitor is being cited and you are not. Five polls across ten prompts beats fifty polls of one. Method for the wider measurement programme is on our page on how to measure share of answer across AI engines; this page is about the narrower comparative case.

How to run a citation-source audit on the answer that named them

Once the gap survives five polls, stop reading the prose of the answer and read its footnotes. A citation-source audit is four steps and can be done by hand in an afternoon.

1. Capture the answer with its sources. Expand every citation chip and record the full URL, not the domain. The page matters: a directory’s category listing and its blog post are different assets.
2. Classify each URL into one of five buckets — a company’s own website, a directory or review platform, news/analyst/research, community and social, or regulator and platform documentation.
3. Mark ownership. Split the company-website bucket into pages the named brand publishes and pages published by anyone else. This split decides your budget.
4. Repeat across five to ten prompts and count buckets, not anecdotes. One audited answer tells you about one answer.

The output is a two-line brief: here is the share of the evidence we could have written, and here is the share we would have had to earn.

What AI engines actually cite on “best provider” questions

We ran exactly that classification over our own monitor. Sample: 1,032 cited sources drawn from 121 answers to 24 commercial and comparison-intent prompts, polled on ChatGPT (browse) and Gemini (grounded) between 1 and 11 September 2026.

Read the first row carefully, because it is the opposite of what it looks like. Company websites dominate — but those 1,032 citations were spread across 541 different domains, 349 of which were cited exactly once, and the ten most-cited domains between them accounted for just 14.2%. There is no single site the engine keeps going back to. For any one brand in the answer, its own pages are a slice of a very wide plate: in the 35 answers in that window that cited LeadsNow, our own domain supplied 46 of the 254 sources — 18.1%. On our own log, publishing more of our own pages moved about a fifth of the input to the answers that named us. That is worth doing and it is not sufficient, which is the honest version of the advice most of this category sells.

Why does AI recommend my competitor? Four causes, and the test for each

Ordered by how often the audit turns them up, not by how bad they are.

Notice that three of these four are diagnosed by looking at other people’s pages. The comparative case is unlike the plain-absence case for exactly that reason: when nobody is cited the fault is usually yours, and when somebody else is cited the evidence is usually sitting on domains you have never logged into.

What closes the gap — and what it honestly costs to run

The work splits cleanly. The half you can do yourself is the on-site half: rewrite the pages you want cited so the answer comes first, add the comparison table, put a real date on them. Our own reading of what separates a retrieved page from a cited one is on how to be in the 15% ChatGPT actually cites, and the per-engine differences are set out in how B2B brands get cited in Perplexity. None of it needs a vendor.

The expensive half is earning placement on the third-party pages the audit found: identifying which cited pages accept submissions or updates, writing something worth publishing, following up. Budget the audit at roughly a day per ten prompts by hand, the monitoring at an hour a week, and the outreach at whatever your team costs — it is the only part that does not compress with tooling. What breaks at volume is the polling: five polls across thirty prompts on two engines is 300 answers a cycle, which is where people quietly stop. Running that programme is what our AI SEO and answer-engine optimisation service does; we point the same monitor at ourselves, and on 11 September 2026 it put us at 1 of 15 prompts on ChatGPT and 1 of 15 on Gemini. We publish that because a supplier who will not show you their own bad week cannot be checked.

What will not close the gap

Three things absorb budget and do not move the audit. First, a machine-readable file: we published llms.txt and measured the result, and whether llms.txt does anything for AI search is a documented negative. Google’s documentation agrees: “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” Second, volume for its own sake — more posts on your own domain compete for the slice of the answer you already control — 18% of it in our own log. Third, asking the assistant why it chose them; models reconstruct a plausible reason after the fact, and that reason is not a log of what was retrieved.

Frequently asked questions

Why does ChatGPT recommend my competitor and not me?

In the audits we run, the most common reason is third-party coverage rather than anything on either website. Across 1,032 cited sources from 121 answers to 24 commercial and comparison-intent prompts polled on ChatGPT and Gemini between 1 and 11 September 2026, the ten most-cited domains between them accounted for just 14.2% of citations, and 349 of the 541 cited domains appeared exactly once. The engine is assembling an answer from a wide, shallow spread of pages, and your competitor is on more of them.

Do I need llms.txt, AI-specific schema or a special file to catch up?

No. Google’s AI Features and Your Website documentation states that “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, and that no new machine-readable files or special schema.org structured data are needed. We tested llms.txt on our own site and published the negative result rather than the theory.

How do I check the engines can even reach my site before blaming content?

Check your server logs and robots.txt for the named crawlers before you conclude anything about quality. OpenAI’s crawler documentation states that OAI-SearchBot is the agent used to “surface websites in search results in ChatGPT’s search features”, and that it is separate from GPTBot, which crawls for model training. Blocking the wrong one is a common own goal, and it is a five-minute check.

Should I get listed on directories and review platforms to close the gap?

Only the specific ones your audit found. Directory, marketplace and review platforms supplied 5.6% of the 1,032 cited sources in our 1–11 September 2026 window across ChatGPT and Gemini — meaningful but small, and heavily category-dependent. Paying to be listed somewhere your audit never surfaced is spending against a guess.

How long should I measure before deciding anything changed?

Months, on a fixed cadence, not weeks. Our own measured rate on ChatGPT moved from 3 of 15 tracked prompts on 9 September 2026 to 1 of 15 on 11 September 2026 with no change to the pages in between. Any programme judged on a two-day swing will be judged wrongly in both directions.

Can I just ask ChatGPT why it picked my competitor?

You can ask, and you should not trust the answer. The model generates an explanation after the fact; it is not reading back a retrieval log. The citation list under the answer is the real evidence, which is why the audit reads URLs rather than reasons.