ChatGPT cites a competitor and not you for a reason you can diagnose: not their website copy or backlink count, but their presence and mention volume across trusted third-party sources, plus content a model can extract cleanly. It is a measurable gap, not magic, and this article shows how to read and close it.
What ChatGPT is actually choosing when it cites a brand
Start with a distinction worth drawing precisely. A mention is your brand named in the prose of an answer. A citation is a linked source the model attributes a specific claim to. Both are counted differently and fixed differently: you can be mentioned without a single link, or a source can be cited while your brand goes unnamed. In practice you can top the mention count and still lose every citation, because the link the model attributes its claim to goes to a review site or a publisher, not to you.
Your standard analytics will not show either. Click and referral-traffic reports register a visit only after someone leaves the answer and lands on your site; they say nothing about whether ChatGPT named or cited you in the reply that visitor never clicked. Citation and mention frequency have to be tracked directly, by reading the answers themselves.
ChatGPT also does not run a live web search on every turn. Many replies are generated from training data, so no fresh source is attributed at all, and the brands that surface are the ones the model already learned. When it does cite, it rarely leans on a single winning URL: Profound's analysis of cited ChatGPT conversations reported around 6 unique citations per conversation on 3 February 2026. A citation gap is a competition for a place in a small set, not a single top spot.
Diagnosing your gap starts the same way every time: run real buyer-style prompts in ChatGPT and read which brands get named and which get cited.
The signals that decide who gets cited
They run from the signal that moves citation most to the signal that moves it least.
Presence on trusted third-party sources
The strongest signal is not on your own domain. It is how often, and how favorably, independent sources name you: review platforms, community forums, publishers. In the reports we generate, the brands cited most often are the ones mentioned broadly across trusted third-party sites, not the ones carrying the largest raw backlink counts. Your own page confirms a claim; someone else's page is what the model reaches for to source it.
A couple of source types stand out. Wikipedia works as a default baseline the model reaches for first — Profound found it in nearly 1 in 6 cited ChatGPT conversations on 3 February 2026. Reddit and community threads carry trust weight out of proportion to their polish, because they read as unincentivized human experience.
Content a model can extract
A source only helps if a model can lift a clean passage from it. Direct-answer openings, short quotable blocks, tables and FAQs raise the odds a passage is pulled intact; a claim buried in a long paragraph is harder to attribute. Review depth is a concrete, measurable input here: Magna's study of AI responses found businesses with strong reviews on three or more platforms such as Google, Trustpilot and G2 were 2.7 times more likely to be recommended than businesses concentrated on a lone platform, in results published 25 February 2026. A thin profile on a single platform is a weak signal.
Comparison and original-data pages
A couple of page types earn citations directly. Dedicated 'X vs Y' comparison pages match a query buyers type verbatim, so a model pulls them when someone asks how you stack up against a rival. Original or proprietary data — a benchmark, a survey, a number nobody else has — gives the model something it can source only from you, which raises citation likelihood. And citations arrive in co-cited packs: your name appears alongside competitors and adjacent trusted domains, not as a lone winner, so the goal is to sit inside the pack the model assembles.
How citation mechanics differ across ChatGPT, Perplexity, Gemini and AI Overviews
The engines do not select or attribute sources the same way, so a separate audit per engine is required. Vendor documentation describes different retrieval designs, and in the reports we generate we see the same buyer prompt return different cited brands and domains across them. The table below summarizes how each retrieves sources and how visibly it attributes them.
| Engine | How it retrieves sources | How visibly it attributes |
|---|---|---|
| ChatGPT | Searches conditionally, often answers from training data | Inline links when it cites, none otherwise |
| Perplexity | Retrieves live sources for most answers by design | Numbered citations shown prominently |
| Gemini | Blends Google's index with model knowledge | Links surfaced selectively |
| AI Overviews | Draws from Google's ranking system on the results page | Named source links inside the overview |
Read the table as a starting map, not a verdict: within each engine the mix shifts by topic and by how commercial the query is, and vendor behavior changes without notice. What holds across all of them in the reports we generate is that a brand absent from trusted third-party sources stays absent no matter which engine you test.
When a citation is wrong, not just missing
Absence is not the only failure mode. A model can name you and get it wrong, and that needs a different fix.
Misattribution is documented. The Tow Center's controlled study of AI search engines, published 6 March 2025, found DeepSeek misattributed the source of supplied excerpts 115 out of 200 times, and the engines it tested repeatedly credited syndicated or outdated copies instead of the original publisher. So a claim that is genuinely yours can be attributed to a competitor who republished it.
Models also tend to re-cite what they have cited before, so a stale source can outlast newer, better content until the newer page accumulates its own third-party signals. Correcting a wrong or negative mention is therefore not the same job as filling an absence: you fix it by getting accurate, well-sourced material onto the trusted third-party pages the model already pulls from, not by editing your own site alone.
A couple of questions practitioners raise deserve plain answers. Licensed content-partner tiers, where a publisher's material is used under a direct agreement, are effectively closed to an ordinary independent brand, so plan around earned citation rather than a partnership. And paid channels — ads, PR spend, sponsored placements — have no measurable direct effect on citation odds in the reports we generate; what they can do is generate the third-party mentions that carry weight.
Where to start closing the gap this week
Treat the first week as measurement, not optimization. You cannot fix a gap you have not sized.
First, write a fixed set of buyer-style prompts — a couple of dozen that mirror how your customers actually ask, including 'X vs Y' phrasings. Run them across ChatGPT, Perplexity, Gemini and AI Overviews. Record, for each, which brands are mentioned, which are cited, and which source the citation points to. Log every observation with the date you ran it, so the sheet becomes a baseline you can re-run and compare against.
Then read the gap for its cause rather than fixing everything at once. If competitors win because they sit on review platforms and community threads you are absent from, that third-party presence is where to start; if they win because their pages answer cleanly and yours bury the answer, extractability comes first. Fix that single signal, keep running the same prompts, and watch the dated sheet. In the reports we generate, citation frequency moves gradually as third-party signals accumulate, not overnight, so the re-run is what tells you it worked.
For the mechanics underneath every engine, see how each AI engine actually chooses its sources.
Run a free Namedrop scan to see which brands ChatGPT, Perplexity, Gemini and Claude cite for your buyer prompts and where your gap sits.