What each engine reads before it answers

"The AI engine" does not exist. There are engines, and they do not behave alike. The one that cites its sources does not pick them the way the one that cites none does, and a company can come up every time on one and never on another.

What follows describes behaviour observed at the time of writing. These products move fast: treat it as a snapshot, not a law.

Perplexity: search first, writing second

The most legible of the five, and by far the easiest to work on.

Perplexity runs a web search on every question, reads the pages that come back, and writes while citing numbered sources inline. You can see where each statement came from. You can click.

Two consequences for you. First, classic search ranking genuinely counts here: being among the top results for a close phrasing noticeably raises your chances of being read. Second, and more useful, the list of cited sources tells you exactly which pages built the answer. That is a map.

When a GEO project starts, this is usually where the picture clears fastest.

ChatGPT: sometimes it searches, sometimes it remembers

The most used, and the hardest to read.

Depending on the question, the model and the configuration, ChatGPT either searches the web or answers from what it learned in training. Both produce an answer that looks the same, and it is not always obvious which happened.

When it remembers, you are facing an old snapshot of the web. Companies that were well known at training time have the advantage, recent ones are absent, and nothing you did last month changes any of it.

When it searches, you are back to Perplexity's logic.

In practice: a poor showing on ChatGPT does not get fixed in one go, and certainly not in a month. This is the engine where you judge a trajectory over two or three quarters, not over one reading.

Google AI Mode: the Google index, summarised

Google composes a written answer from its own index, with links to the pages it used.

This is where classic search work transfers most directly. If you already rank well, you have a real chance of being picked up. If you are nowhere, you will not appear.

It is also the one that worries traffic-dependent sites most, for an obvious reason: the answer sits above the links, and plenty of people stop there. Being cited without getting the click becomes a result in itself, and that forces a change in what you measure.

Gemini and Claude: the particular cases

Gemini leans on the Google ecosystem and can fetch the web. Claude answers mainly from what it learned, with search depending on the usage context.

In both cases, consumer volume for "which supplier should I pick" questions is lower than on the first three. They matter most if your customers are technical, or if the people you sell to already work with these tools daily. Which is increasingly common in B2B.

I would not make them a priority for a local business. I would watch them if you sell to product teams, developers, or professional firms.

What to do with all this

Three plain rules come out of it.

Do not measure a single engine. The gaps between them are large, and a company well cited on Perplexity can be invisible on ChatGPT. One blended number averages realities that have nothing to do with each other.

Start the work with the engines that search. Perplexity and Google AI Mode react to what you do, within a reasonable delay. ChatGPT follows, more slowly, because what you build today becomes material for its next training run.

And keep every answer exactly as it came, with its date and the engine named. Without that, in three months you will no longer know what changed in the world and what changed because of you.