Perplexity cited our brand as the strongest engine in a six-model test. In the same test, on the same prompt slice, ChatGPT cited us zero times across eight tracked prompts. Same site, same pages, same day of measurement, opposite results. That is not a content-quality problem, because the same Geoboard baseline ranked the same domain first in five of six models on a separate identity prompt using the same underlying pages. The split comes from how each product retrieves and grounds an answer, not from what your page says. This article explains the mechanism behind that split, shows the first-party numbers behind it, and sets out what to actually do when one major assistant cites you and the other one does not, which is a more common and more instructive pattern than most AI-visibility reporting admits.
If you have already read our companion piece on AI-citation monitoring cadence and metrics, this article answers a different question. That guide is about how often to check and what to log. This one is about why the two most-cited Western assistants disagree so often on the same brand, and what that disagreement should change about how you read a report, brief an agency, or sign off on a GEO vendor.
What the Perplexity/ChatGPT split looks like in practice
The pattern is not subtle once you separate the engines instead of averaging them. Across our own tracked prompt set, the two products behave differently enough that reporting one number for “AI visibility” hides the more useful story.
| Behaviour | Perplexity | ChatGPT (default consumer product, same period) |
|---|---|---|
| Default retrieval pattern | Live web search on most queries, by design | Narrower browsing invocation in typical consumer sessions |
| Our Geoboard baseline (2026-06-11), tracked slice | Strongest of six models tested | Zero presence across eight tracked prompts |
| Identity prompt result (same baseline, same pages) | Contributed to a #1-in-5-of-6 result | Contributed to the same #1-in-5-of-6 result |
| Sensitivity to page freshness | High; favours recently updated, narrowly indexed pages | Lower in the tracked slice; leans on broader association |
| HTML vs JavaScript rendering (CitationOne/Alpar, June 2026) | Reads raw HTML only | Reads raw HTML only |
| Typical citation format | Named source link alongside the answer | Named brand mention more often than a linked citation, in our own logs |
Two things in that table matter more than the headline gap. First, the identity-prompt row is identical for both engines, which means the ChatGPT zero is not a general inability to find or describe the brand. It is concentrated in a specific query family, the same transactional and informational slice where our earlier baseline article documented a broader zero. Second, the rendering row is identical too, which rules out the most common technical excuse. Whatever explains the gap, it is not that ChatGPT cannot parse the page and Perplexity can.
Why the two engines ground answers differently
The honest answer stays close to observed behaviour and avoids guessing at internal architecture neither of us can see. What we can describe is the pattern, repeated across our own measurement cycles: Perplexity is built and marketed as a search product, and its default behaviour on most factual or comparative queries is to run a live web search and cite what it finds, even for narrow or recently updated pages that have not yet accumulated much third-party authority. That is the product’s core value proposition, so live retrieval is the rule rather than the exception.
A general-purpose assistant’s default consumer behaviour, in the same measurement period, leaned differently. Not because it is incapable of browsing, but because the query itself, the session context, or the product’s own routing decided a live fetch was not the default path for that particular prompt. The practical effect is the same regardless of the exact cause: a page that Perplexity would fetch and cite today may never get looked up at all in an equivalent ChatGPT session, which means the citation opportunity never opens in the first place.
This is a retrieval-invocation difference, not a comprehension or rendering difference. The CitationOne/Alpar test from June 2026 confirms the second half of that claim directly: both engines read raw HTML only when they do fetch a page, with no meaningful gap in how well they parse what they find. The gap sits earlier in the pipeline, in whether a fetch happens at all for a given query, which is exactly why averaging the two engines into one visibility score destroys the signal that would tell you where to spend remediation effort.
One more mechanical point worth stating plainly: live search grounding rewards freshness and narrow indexability in a way that trained-knowledge grounding does not. A page updated last month describing a specific service, with clear entity signals and no ambiguity about who offers it, is exactly the kind of result a live-search product surfaces well. The same page, if the assistant answering the question never triggers a fetch for that query, gets no chance to be evaluated at all. That is a retrieval-opportunity gap, and it explains our data better than any theory about content quality.
First-party proof from our Geoboard baseline
We do not publish this claim as a general theory picked up from someone else’s audit. We measured it against our own site. Our Geoboard baseline, dated 2026-06-11 and described in full in how we tracked our own AI citations for 90 days, ran the same fixed prompt set against six models. Two results from that single run matter for this article specifically.
The first result: on the narrow identity prompt “Polish agency for foreign WordPress clients,” wppoland.com ranked first in five of six models. The second result, from the same run, same date, same underlying pages: ChatGPT showed zero presence across eight tracked prompts in the blind-spot slice of that prompt set, while Perplexity was the strongest of the six engines tested. Both results came from the same batch, so neither one can be explained by a difference in content freshness, page changes, or measurement timing between them. The variable that changed was the query family and the engine, not the site.
That combination is the actual finding worth taking seriously, more than either number alone. A brand can be correctly identified and even ranked first by an engine on one query type while being invisible to a different engine on a different query type, using the exact same web presence. Any monitoring approach that reports a single aggregate score across engines and query families cannot represent that reality. It will either flatten the ChatGPT zero into a passable-looking average, or, less charitably, it will hide the zero behind a Perplexity win that a client or stakeholder never gets to see broken out.
When API monitoring lies to you
The engine split gets worse, not better, once you add measurement method into the picture. Our Q2 2026 measurement cycle, documented in full in measuring our own AI citations, ran three snapshots across a quarter and exposed a second axis of disagreement that sits underneath the engine split: API proxy results and real consumer-product results do not agree, and the gap between them can be larger than the gap between engines.
In April, an API-proxy snapshot reported a 7.7 percent brand-mention rate and a 0 percent URL-citation rate for ChatGPT across 26 queries, with directories and job listing sites cited instead of the brand’s own pages. In May, a second proxy snapshot returned an undetermined result for all 20 queries tested, which is the instrument admitting it could not tell whether grounding had happened at all. Only in June, when the same team switched to reading real consumer-product outputs instead of an API proxy, did the picture line up with what the later Geoboard baseline would confirm: ChatGPT weakest, Perplexity strongest, on the same brand.
The lesson generalises past our own site. An API proxy calls a model endpoint and inspects the raw text response. That is cheap, repeatable, and easy to script into a dashboard. It is also frequently not the same product a real customer is using, and the two can disagree enough that quoting a proxy number as “AI visibility” without stating the method is close to meaningless. If a report claims a ChatGPT citation rate, or a Perplexity citation rate, the first question to ask is whether the number came from the actual consumer interface or from an API call standing in for it. Our own numbers moved more between measurement methods in the same quarter than they later moved between the two most relevant engines. That should worry anyone buying a monitoring report on the strength of a single clean-looking chart.
Who should care about this split
This is not a niche concern for SEO practitioners alone. Three groups make decisions that the Perplexity/ChatGPT split directly affects, and each one is currently more likely to see a blended number than the split that would actually inform their decision.
B2B technical evaluators researching a vendor increasingly ask an assistant before opening a search engine at all. If that evaluator defaults to ChatGPT and your brand’s citation behaviour is concentrated in Perplexity, you are invisible to a meaningful share of your own research funnel, and no amount of on-page polish changes that until the retrieval-opportunity gap itself is addressed.
Procurement teams evaluating a GEO or AI-visibility vendor are the audience most exposed to the averaging problem, because a vendor selling a monitoring product has every incentive to report one reassuring number rather than two numbers where one is uncomfortable. Procurement should treat any AI-visibility figure that does not name the engine, the exact prompt, the date, and the method as unverifiable, for the same reason it would reject an unsourced financial claim.
Agencies reporting AI-citation results to their own clients carry the direct professional risk here. Reporting a single “AI visibility improved” line when the underlying reality is “Perplexity improved, ChatGPT stayed at zero” is the kind of claim that looks fine in a slide and falls apart the first time a client asks a follow-up question. The honest version of that report is longer and less flattering, and it is also the version that survives scrutiny.
What to fix when Perplexity cites you but ChatGPT does not
This is the common direction of the gap in our own data, and probably in most brands still building AI-specific authority rather than general-web authority. Three things are worth checking, in order.
First, confirm the gap is real and not a measurement artifact. Re-run the same exact prompts, not paraphrases, in the actual ChatGPT consumer product, not an API proxy, and log the date, the prompt wording, and the outcome. Our own Q2 cycle showed how much a proxy snapshot alone can mislead.
Second, separate the query families before assuming the whole brand is invisible to ChatGPT. Our baseline showed the gap concentrated in transactional and informational query families, while an identity prompt using the same pages still returned a strong result. A brand-wide ChatGPT problem and a query-family-specific ChatGPT problem call for different remediation, and conflating them wastes effort on the wrong fix.
Third, and this is the fix that actually moves the needle for the transactional gap specifically, invest in off-page corroboration rather than another round of on-page content. A live-search product like Perplexity rewards a well-structured, current page directly. A general-purpose assistant leaning on broader association rewards a brand that the rest of the open web already corroborates, third-party mentions, independent technical answers, verifiable directory listings, and named contributions in places the assistant’s training or retrieval already trusts. On-page work is the floor. It is not the lever that closes this particular gap.
What to fix when the reverse happens
The reverse pattern, ChatGPT citing a brand that Perplexity does not, shows up less often in query families driven by freshness or narrow niche matching, but it is worth a short note because the fix looks different. If a brand has strong general-web association, backlinks, press mentions, a well-known name, but its own pages are thin, outdated, or slow to update, a general-purpose assistant leaning on broader trained association can still surface the brand from memory or wide corroboration, while a live-search product fetching the actual current page finds less to cite there and moves on to a competitor’s fresher result.
The fix in that direction is almost the mirror image of the first case: less about off-page authority, which is already present, and more about making sure the live, fetchable version of the page actually reflects the authority the brand has elsewhere. Update the page, make the specific claims a live-search fetch would want to cite explicit and current, and confirm the content that supports the brand’s reputation is not sitting only in press coverage or third-party profiles that a live crawl would not associate directly with the domain being queried. This direction of the gap is rarer in our own measurement, but it is the one most likely to blindside a brand that assumed general reputation would automatically transfer to live-search citation behaviour.
How to monitor both without averaging
None of the above is actionable without a monitoring approach that keeps the two engines, and the two measurement methods, separate at every step. Our companion guide on AI-citation monitoring cadence and metrics covers the full cadence table and metric list in detail; the summary relevant here is narrower.
Log every result against four fields: the engine, the exact prompt, the date, and whether the source was the consumer product or an API proxy. Never collapse Perplexity and ChatGPT results into one average, and never quote a proxy number without stating that it came from a proxy. Re-run a fixed prompt list weekly in the real consumer interfaces to catch product changes fast, and reserve heavier multi-model batch tooling, the kind that produced our own Geoboard baseline, for a six-to-eight-week cadence where trend comparison actually means something.
If you are building or auditing this kind of monitoring for the first time, start by confirming your site’s content is even eligible to be cited at all. Most Western AI assistants read raw HTML only, not executed JavaScript, which is a gating precondition that sits underneath every engine-specific result described in this article. A zero caused by client-side rendering looks identical to a zero caused by a genuine retrieval gap until you check which one you are actually looking at.
Closing note
The single most useful thing our own data has produced so far is not a score. It is a split: one engine that cites us and one that, on the same day, on the same pages, did not. That split is not a defect to hide from a client or a stakeholder. It is the most actionable piece of information a monitoring programme can produce, because it tells you exactly which mechanism to fix, off-page corroboration for the ChatGPT-shaped gap, page freshness and specificity for the rarer reverse case, rather than sending you back to rewrite a page that was never the actual problem. We are tracking whether this split narrows over the next ninety days as part of our own 90-day AI-citation series, and we will report it the same way we are reporting it here: by engine, by date, and by method, never averaged into a number that flatters one result and hides the other. If you want this diagnosis run against your own property, with the split broken out by engine and query family rather than compressed into one score, that is the starting point of our AI and LLM visibility playbook and the programme our GEO and LLMO optimization service delivers.
