Which AI Assistant Cites You? Testing Perplexity, ChatGPT and Claude

If you ask three different AI assistants to recommend a project management tool for a five-person startup, you will get three different answers with three different kinds of sources — even on the same day, with the same web. That is not a glitch. The engines find, select and attribute products through genuinely different mechanisms, and optimising for one tells you nothing about the other two.

This post is the companion to our AI crawlers guide and the GEO primer. Those cover how to be reachable. This one covers how to be cited — per engine — and how to test yourself without buying anything.

Why one "AI SEO" strategy cannot work

The three assistants answer "what should I use?" through different pipelines:

• Perplexity runs a live search on nearly every product question and cites its sources inline. The answer is assembled from pages it fetched seconds ago. Freshness and page-level structure dominate.
• ChatGPT blends a trained model with an optional search mode. When search is off (or the user is on a plan without browsing), the answer comes from training data — which means it comes from whatever was written about your product months ago, not from your own site at all.
• Claude leans hardest on training data and is the most conservative about naming specific vendors. When it does cite, the citation usually traces back to a third-party description of your product — a directory listing, a comparison article, a review — rather than your homepage.

The practical consequence: the same product can be perfectly cited in Perplexity, vaguely described in ChatGPT, and invisible in Claude. That is three different problems, not one.

A test protocol you can run for free

You do not need a GEO tool to find out where you stand. You need discipline about the variables.

Build a fixed question set. Write 8–10 questions a real buyer would ask, mixing three shapes:

• category questions — "what are good AI meeting notetakers for remote teams?"
• product questions — "is [your product] good for agencies?"
• comparison questions — "[your product] vs [competitor]"

Control the conditions. Fresh session for every run (logged out where possible), same day, note the date. Run each question in each assistant. Record three things per run: were you mentioned, were you linked, and which sources were cited instead of you.

Repeat across days. Search-backed engines rotate sources between runs. One run is an anecdote; five runs across a week is a signal. A spreadsheet is fine. What you are looking for is a stable pattern, not a precise number.

Watch your analytics in parallel. Referrers like perplexity.ai and chatgpt.com show up as normal traffic sources. In our own logs the volume from assistants is small but growing, and — more usefully — the pages they fetch show up in crawler logs, which is how we confirmed our llms.txt was being pulled at all.

What each engine actually rewards

Perplexity: be fetchable, be fresh, be structured

Perplexity cites what it can retrieve and read right now. The levers, in order:

1. Let the right bots in. PerplexityBot (crawling) and Perplexity-User (user-triggered fetches) must be allowed in robots.txt. Blocking the first means you never enter the index; blocking the second means a user who explicitly asks about you gets a stale summary.
2. Publish facts in a machine-friendly shape. Lists, tables, explicit comparison rows, and a page whose first paragraph states what the product is and who it is for. Perplexity's citations often quote a single paragraph — write paragraphs that can survive being quoted alone.
3. Stay fresh. New content and updated dates move citations between runs in a way we did not observe in the other two engines. A directory listing that was updated this month out-cites a homepage last touched two quarters ago.

ChatGPT: the answer often comes from someone else's page about you

With browsing off, ChatGPT describes your product from training data — aggregated from articles, directories, changelogs and discussions. You cannot edit the model. You can edit the corpus it learned from:

• Make sure at least one authoritative third-party page describes your product in the terms you want used. A complete directory listing does exactly this, and it is why we seed every launch on aat.ee with a plain-language description.
• Prefer phrasings that are easy to summarise: what it is, who it is for, what makes it different. Models reproduce clean claims better than clever ones.
• When browsing is on, OAI-SearchBot needs to be allowed — it is the fetcher behind the linked answers.

Claude: third parties do the talking

Claude names vendors sparingly and hedges visibly. When it does commit, the description almost always paraphrases a source other than your own marketing site — a review, a directory, a comparison. The optimisation is unglamorous:

• Get described accurately in enough independent places that the paraphrases agree.
• Keep those descriptions current. A renamed or repositioned product with stale third-party copy will be described as its old self for months.
• There is no Claude crawler to allow today; there is only the corpus. That makes Claude the engine where directories and structured listings carry the most weight per hour invested.

The fixes, side by side

One row deserves emphasis: third-party descriptions are the only lever that moves all three engines. It is also the lever most founders skip, because it feels passive. It should not be — a listing is a page you control that lives on someone else's authority. That is the entire premise of what we built aat.ee around.

Run the test before you buy anything

The GEO tool market is happy to sell you a score. Scores are fine, but the underlying question is simple enough to answer yourself in a week: when a buyer asks an assistant for a product like mine, am I named, am I linked, and is the description accurate?

Run the protocol. Fix the engine-specific lever that is weakest. Then run it again next month — the engines change under you more often than Google does.