The state of AEO in 2026: Claude is not ChatGPT

We enabled web search on both models and tested prompts that ranged from current product recommendations to basic explainers. Claude searched 36.6 percent of the time. For the other 63.4 percent, any search optimization work was outside the path because the model answered without retrieving the web.

The language of the prompt changed the route. Terms such as "best," "near me," and a current year tended to trigger retrieval. Basic "what is" and "how does" prompts were more likely to stay inside the model, depending on the subject and freshness needed. Search invocation is the first gate, and that gate changes by prompt class.

Claude's citations followed Brave closely. We found 79.2 percent of cited URLs within Brave positions one through ten. The ordering was largely preserved rather than rebuilt through a separate visible ranking layer.

Claude gives us a route we can inspect. Check whether the prompt triggers search, inspect the fanout, then run those strings through Brave. A page missing from Brave's first page is unlikely to enter Claude's citation set for that path.

Claude's source selection looks different from ChatGPT because the retrieval path is different before either model writes a sentence.

Claude's cited domains overlapped 64 percent with Google results. ChatGPT's overlap was 37 percent. Yet Claude and ChatGPT shared only 8 percent of citation domains on average.

There is no universal AEO result page hiding behind the interfaces. A page can be visible to Claude through Brave, visible to ChatGPT through a different fanout, and absent from another engine that never searched for that prompt.

Claude cited listicles in 36.4 percent of the classified set, compared with 19.7 percent for ChatGPT. ChatGPT used forums and user-generated content in 15.8 percent. Claude used them in 0.9 percent, a gap of more than seventeen times.

The listicle finding fits the Brave dependency. Search results often reward pages that gather options under a current, query-shaped title. ChatGPT's larger forum share produces a different evidence base, with more first-person experience and disagreement.

The same query strings appeared in Claude's fanout roughly 65 percent of the time. It added "2026" or "2025" to 94 percent of fanouts, compared with 17 percent for ChatGPT. Adding the year to a title made it 17 percent more similar to the queries Claude generated in the measured set.

Use a year when the answer actually changes with time, then update the evidence. Claude's behavior rewards recency language because it generates recency-shaped searches.

In May 2026, we also saw a sharp change in ChatGPT referral traffic. It rose 60 percent overnight in the observed dataset and settled at roughly 1.6 times the prior global level. One in four clicks landed on a homepage. Brands were being hyperlinked more often inside answers, which gave homepages a role that earlier citation studies rarely showed.

Ads were moving into the same conversation. The observed product used prompt context to match ad titles and descriptions, with roughly one ad per minute and one per conversation in the tested experience. I expect placements to move deeper into threads as inventory grows, with tighter matching to the conversation.

Google AI Mode can answer from Google's first-party place and product systems or search third-party pages. A prompt such as "brunch near me" can stay close to Google's own data. A prompt asking whether a new phone is worth buying needs evidence from outside pages.

That split changes the work. Local and inventory questions depend on complete entity data. Evaluative questions depend on clear third-party sentences that name the brand, product, place, and claim without relying on a pronoun two paragraphs away.

My first step is to identify whether the engine searches, then trace the index, fanout, and source type for that prompt. Claude and ChatGPT share only 8 percent of citation domains on average, so one proxy metric cannot tell you where a page stands.