Should you optimize content differently for ChatGPT, Gemini and Perplexity? What we tell clients

Clients ask this in the second week of nearly every engagement, usually after someone read a post claiming that each engine wants something different. Our answer has been stable for a while: write one strong page, make sure every engine can reach it, and measure what each engine does with it. Engine-specific rewrites come last, and only when the data asks for them.

One strategy, because the fundamentals overlap

Google's documentation on AI features in Search says plainly that the best practices for SEO remain relevant for AI Overviews and AI Mode, that there are no extra requirements, and that no special optimization or new schema is needed. A page must be indexed and eligible to show with a snippet. The advice it gives is the old advice: allow crawling at robots.txt and CDN level, link pages internally, keep important content in text, make structured data match the visible page.

For ChatGPT and Perplexity there is no equivalent single document we can point to that rules out engine-specific tricks, so we rely on what we can test. The best published research we know of is the 2024 GEO paper from Princeton and collaborators (KDD 2024). It found that adding citations, quotations from relevant sources and statistics raised visibility in generated answers by up to about 40% on its benchmark, with gains up to 37% on Perplexity, and that the effect varied by domain. We read that as: substantive, sourced writing helps, and your category may behave differently, so test.

What actually differs by engine

Three things, none of them about prose style.

Which bots reach you. Engines fetch through different crawlers. ChatGPT has GPTBot, OAI-SearchBot and ChatGPT-User. Perplexity has PerplexityBot and Perplexity-User. We have seen sites that allow one family and block another through a firewall rule nobody remembered, which produces the odd result of being cited in one engine and absent from the next.

Which of your pages get chosen. Engines run different searches, called query fan-outs, so they retrieve different pages. One may cite your documentation for a prompt where another cites a third-party review.

Which outside sources they trust for your category. This is the big one. If two engines cite different review sites for the same question, your work in each is different, even if your own page is identical.

How we test it

Fixed prompt set. Twenty to forty prompts per client, covering category, comparison and brand questions.
Same prompts, every engine, daily. We compare visibility, position and cited sources per model, not blended.
Crawler check. We confirm which bots fetch which pages and whether errors are occurring.
Per-engine gaps. Where visibility is high in one engine and low in another, we compare the cited sources and the fetched pages.
One change at a time. Then we re-measure.

The tooling

We run client programs on Promptwatch because it separates results by model. Monitors run per model, country and language, and a dashboard template gives an AI model breakdown. A heatmap can compare any two dimensions, such as models against topics. Citation analytics shows the domains and pages each engine cites, and query fan-outs per model show how each one searches. Agent Analytics logs IP-verified AI crawler visits by bot name, with a crawl-to-citation path per page, so a blocked bot shows up as a failed or missing visit rather than a mystery.

Real-interface tracking covers ChatGPT search, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Copilot, Meta AI and Alexa. Models like Claude, DeepSeek and Mistral are tracked via API alongside, which we note in client reports so nobody mistakes an API sample for what a user sees in the app.

For plans, Essential ($95 a month) covers the model-level tracking and citation reports but not crawler logs. Professional ($245) adds crawler logs. Business ($579) adds the Content Agent and local targeting. Agency self-serve plans ($199, $399, $799) include all models and crawler logs. A free Explore plan covers ChatGPT only with 10 prompts.

What we do not do

We do not add markup because a post says one engine loves it. Google states it does not need special schema for its AI features, and we hold other claims to the same standard: show us the data. We do not promise a client that a change will win a particular engine, because engines vary between runs and update often.

A practical order of work

First, check crawl access for every relevant bot. Second, tighten the pages that matter: a direct answer near the top, sourced facts, clear headings. Third, close gaps in third-party coverage where one engine leans on sources the client is missing from. Only then consider engine-specific adjustments, and treat each as an experiment with a before and after.