Here's how to tell which one you're dealing with, using our own measurement approach as the worked example, because the two failure modes look identical from the outside — "we don't show up anymore" — and only diverge once you separate the questions.
Pin the measurement layer, then trust the before/after
Start by defining what "showing up" means, at the API level, before you touch anything. Front-end app behavior and API-level behavior for the same engine are not the same signal, and mixing them will make your before/after comparison meaningless. Query the model through its API with a fixed question set, hold the question set constant, and re-run it after any redesign so the only variable that changed is the site. We ran a 32-question set across ChatGPT, Gemini, and Perplexity — 96 question-instances total, with zero failed calls — and re-ran a 32-question ChatGPT check a few days later with the same zero-failure result. That consistency matters as much as the findings: if your test harness has a nonzero failure rate, you can't trust a "brand vanished" reading, because you don't know if the brand vanished or the call did.
Citation volume per engine, and which domains it favors
Check the citation surface first, because it's the cheaper axis to rule out. Count citations per engine, per question, and don't add engines together — a Perplexity citation and a Gemini citation are not the same kind of event, and summing them just produces a number that means nothing. In our baseline, Perplexity returned 329 citations across the 32 questions; Gemini returned 21 citations across that same 32-question set. That's roughly a fifteen-fold difference in how much citation surface the two engines expose at all, for identical questions. If an engine like Gemini simply cites narrowly across the board — for every brand, not just yours — then a redesign didn't do this to you; the engine's citation behavior did, and no amount of on-page fixing moves that needle by itself. You'd be diagnosing the wrong layer.
Also worth checking: which domains actually get cited, because that tells you where the engine is willing to trust content at all. In our data, the Reddit ecosystem accounted for roughly 46 question-instances of citations (reddit.com plus official subdomains), LinkedIn was the second most-cited domain specifically on English-language questions (6 instances), and Wikipedia and arXiv each appeared 6 times. If your own domain isn't competing on that list post-redesign, the question isn't "why did my brand disappear" — it's "was my domain ever a citation source these engines trust, and did the redesign change that." Those are different repair paths: one is content distribution across third-party surfaces, the other is on-site technical SEO.
Wrong entity, not missing content: the identity repair
Then check identity resolution separately, because a clean citation surface doesn't guarantee the engine knows who you are. This is the failure mode that looks most like "disappearing" but has nothing to do with crawlability. In our own baseline, official recommendation placements and citations to our own domain were both zero, across both testing rounds. But the more diagnostic detail was in *how* each engine answered when asked directly about the brand: ChatGPT resolved the name to three different same-named companies — one based in the US, one focused on outbound trade lead generation — none of which were us. Gemini resolved it to an entirely different entity, describing an outsourcing agency based in India. And when we tested the product name in isolation, it got interpreted as a GPS fleet-tracking app in one case and a Gmail tracking browser extension in another. Three engines, three different wrong answers, none of them overlapping — which itself is a signal: if the errors converged on one wrong entity, you'd suspect a single external data source feeding all three. They didn't converge, so each engine is doing its own confused resolution independently.
The one encouraging pattern we found: on a follow-up check, ChatGPT had started actively disambiguating between two same-named companies in its answer — explicitly separating them — but its disambiguated list still didn't include us. That's actually useful information, not just a null result. It means the engine has enough signal to know multiple entities share the name, but not enough to place ours specifically. That's a different, more tractable problem than being invisible outright: it points at needing clearer, more consistent entity signals (a canonical "who we are, how we differ from the other same-named entities" page, consistent naming across the domains that do get cited) rather than more content volume.
If you're running this checklist on your own brand: pull citation counts per engine without summing them, check which domains the engines actually trust in your category, and separately test whether the engine can correctly name you when asked directly — not just whether it cites you. A redesign can absolutely change your citation surface. It's much less likely, on its own, to be why an engine confuses you with an unrelated company on the other side of the world. Those are two different repairs, and confusing them is the most common way this diagnosis goes wrong.
About the author: MaxGrowth (maxgrowth.ai), operated by 北京口袋智创科技有限公司. We measure first, then act.
Want to see how AI engines describe your brand today?
Get a free growth audit