Perplexity runs a live web search for practically every question and writes its answer from the pages it just retrieved. ChatGPT answers from what the model already contains, and searches only when something decides a search is warranted. That one difference in architecture explains almost everything else about how the two behave, including why a site can be cited constantly by one and never named by the other.
It is worth being precise about this, because the industry talks about AI search as though it were one thing with one set of rules. It is not. There are at least two designs in wide use and they reward different work.
Search-first and memory-first
A search-first system treats retrieval as mandatory. The question arrives, a search runs, documents come back, and the model writes from those documents. If nothing is retrieved, the system has very little to say. Perplexity is built this way and makes no secret of it. Its developer documentation describes grounding answers in live web results as the product rather than as a feature bolted onto one.
A memory-first system treats retrieval as optional. The model has already absorbed an enormous amount about the world and can answer a great many questions without looking anything up. A search runs when the question looks like it needs current information, or names something obscure, or the user asks for it. Otherwise the answer comes from parameters.
Neither design is better. They fail differently, which matters more.
| Search-first (Perplexity) | Memory-first (ChatGPT by default) | |
|---|---|---|
| Retrieves on | Nearly every query | Queries judged to need it |
| Answer is built from | Pages fetched seconds ago | Training data, plus pages when it searched |
| Citation density | Most claims carry a numbered source | Sources appear when a search happened |
| Favours | Recent, crawlable, clearly written pages | Well represented, widely repeated entities |
| Your fastest lever | Crawlability and freshness | Corroboration across independent sources |
| Typical failure | Cites a weak page confidently | Omits you entirely without explanation |
| How fast you can move it | Weeks | Quarters |
Why this is the most useful distinction in AEO. If you appear in Perplexity and not in ChatGPT, your pages are fine and your entity is weak. If you appear in neither, start with crawler access. Those are opposite diagnoses with opposite first moves, and one test tells them apart.
Where Perplexity’s sources actually come from
Perplexity operates its own crawler and has historically supplemented its own index with third party search infrastructure. That combination matters to you for a practical reason. Independent indexes exist beyond Google and Bing, and they have different coverage. Independently built web indexes are not a hypothetical any more, and a site that is well covered by one may be thinly covered by another.
The operational consequence is unglamorous and specific. Check what your robots.txt and your firewall do with crawlers that are not Googlebot. We find blocked or rate limited AI crawlers on roughly a third of the sites we look at, and in almost every case nobody chose it. It arrived with a bot management rule, a security plugin default, or a well meaning attempt to reduce server load.
That check takes ten minutes and is the single highest yield thing on this page. A site that blocks the crawler cannot be retrieved, and a site that cannot be retrieved cannot be cited, which is the distinction we worked through in indexed, retrieved, cited.
Citation density, and what it does to your odds
Perplexity attaches numbered sources to individual claims rather than listing a few references at the end. That changes the shape of the opportunity in a way people underestimate.
When an answer carries fifteen numbered citations, there are fifteen slots rather than one. You do not have to be the best source on the topic. You have to be the best source for one specific claim inside a longer answer, which is a far easier thing to be. A page that owns one sharp, checkable fact will be cited next to pages from organisations a thousand times its size.
This rewards a particular kind of writing. One idea per paragraph, the claim stated plainly, the number in the sentence rather than in a chart. It is the same structural property that decides retrieval generally, covered in how retrieval decides which sources to pull, and search-first systems simply apply it more visibly.
The uncomfortable part, which is accuracy
Dense citation looks like rigour. It is not the same thing, and the research on this is not flattering to any of these products.
The Tow Center for Digital Journalism at Columbia tested eight AI search tools on their ability to identify the source of excerpts from news articles. They found the tools were confidently wrong at a high rate, rarely declining to answer, and in some cases producing URLs that did not exist. Premium tiers were not reliably better and were sometimes worse, because they answered more often rather than more accurately.
Two things follow from that, and they pull in opposite directions.
The first is that a citation is not a verdict on quality. Being cited means your passage was retrieved and looked useful for a claim. It does not mean a system judged you authoritative, and you should not report it internally as though it did.
The second is that misattribution is a real risk to you rather than an abstract concern about media. If an engine will invent a plausible URL, it will also attribute a competitor’s claim to your domain, or yours to theirs. The only defence is being unambiguous about what you actually say, which comes back to entity resolution rather than to content. What that means in practice is set out in what an entity is in AI search and in what a knowledge graph holds about you.
The check almost nobody runs
Ask each engine for its source on a claim about your company, then open the link. A meaningful share of the time it will point at a page that does not contain the claim, or at a page that no longer exists. That is worth knowing before you build a reporting dashboard on top of citation counts.
If you would rather have that done at a proper sample size with the transcripts attached, it is part of our free AI visibility check.
What actually differs in what you do
Less than the framing suggests, which is good news. The work overlaps heavily. What differs is the order and how quickly you see anything.
| If your gap is in | Do this first | Expect movement in |
|---|---|---|
| Perplexity | Crawler access, then server-rendered text, then freshness dates | Two to six weeks |
| ChatGPT with search on | Self-contained passages on the pages you want quoted | Four to ten weeks |
| ChatGPT without search | Third-party corroboration and consistent public records | One to three quarters |
| Google AI Overviews | Conventional indexing health, then passage structure | Weeks to months |
| All four at once | Crawler access. It is nearly always crawler access | Days to see, weeks to compound |
The memory-first case is the slow one because you are not trying to win a retrieval slot, you are trying to change what a model absorbed about your category from the wider web. That happens through other people publishing about you consistently, and it does not respond to anything you do on your own domain in under a quarter. The mechanics of that are in how ChatGPT picks which companies to name.
A caution about optimising for the loudest engine
Perplexity is heavily discussed relative to how many people use it. Audience research consistently finds that using AI chatbots for information gathering is still a minority behaviour overall, concentrated among younger users and growing quickly, which is roughly what the Reuters Institute finds year after year on news consumption. Your category may be far ahead of that average or far behind it.
So measure your own. If your buyers are engineers, assume adoption is well ahead of the general figure. If they are procurement officers in local government, assume it is behind. Building a programme around an engine your buyers do not use is an expensive way to be early.
The defensible position is that the underlying work pays off regardless. Crawlable pages, plain passages, consistent public records and current pricing are not Perplexity tactics. They are the conditions for being usable by any retrieval system, including ones that do not exist yet, which is most of the argument in what AEO actually is.
What we would check this week
Ask both engines the same five questions about your company, logged out. Note which claims carry a citation and open every one. Then check your server logs for PerplexityBot, OAI-SearchBot and ChatGPT-User over the last month.
Three outcomes. Crawlers absent from the logs means you have an access problem and everything else is premature. Crawlers present but citations pointing at other people’s pages means you have a passage problem. Crawlers present, your pages cited, and the claims slightly wrong means you have a freshness problem, which is the most fixable of the three and the one that damages deals fastest.
Related reading
- Why AI Overviews cite some brands. The third architecture, which behaves like neither of the two here.
- Mention, citation, recommendation. Three outcomes people report as one number, with different causes.
- The 20-prompt set. The structured way to run the comparison above across four engines rather than casually.
Run the same five questions through both and compare the citations, not the prose.
The prose will sound similar. The citations will not, and the difference between them is the most informative thing you can look at this month. It tells you, for free, whether your problem lives on your servers or out in the wider web.
When the split is clear and the fix is not, send us the domain and we will run a visibility check. You get the transcripts, the crawler log read, and one named constraint rather than a list.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.