Ask most content teams what an AI content strategy looks like for the second half of 2026 and you'll get a version of the same plan: refresh the pages that matter, hit publish, wait for the citation to show up. That plan works reasonably well for some engines. It does not work the way most teams assume for Claude, and a new dataset from Profound's Josh Blyskal spells out exactly why. Claude spends most of its time answering from what it already knows. Not from what it just read.
Claude barely searches the web. Your AI content strategy needs to account for that.
Blyskal runs AEO research at Profound, and his July 22 piece, "The state of AEO in 2026: Claude is not ChatGPT," tested a large set of prompts against Claude and logged whether the model actually reached out to the live web or answered from what was already sitting in its weights. The number that should reorder a lot of content calendars: Claude invoked live web search for only 36.6% of the prompts tested. Just over a third. The rest of the time — nearly two out of every three prompts — it answered from training-time knowledge, no fresh fetch involved.
That's a very different retrieval posture than most GEO advice assumes. Most Claude AI search guidance published this year still treats every engine as a live-lookup machine: crawl the page, rank it, cite it, repeat. Blyskal's data says Claude mostly isn't doing that lookup at all. When it does search — the 36.6% slice — the citations it pulls back lean heavily on one source: 79.2% of them traced to Brave's top 10 organic results. So the fraction of the time Claude is browsing, ordinary search ranking still matters quite a bit. The rest of the time, ranking doesn't enter the picture, because there's no search happening to rank into.
The 63.4% figure is the derived one — it's just what's left after subtracting the search-invocation rate from the full prompt set — but it's the number that should reset how a content team plans around Claude specifically. If most Claude answers come from memory rather than a fresh page load, then a page's publish or update date is, for most of that traffic, irrelevant to whether Claude cites it. What matters instead is whether the page (or the claims on it) made it into what Claude learned in the first place, and whether enough of the rest of the web agrees with it that the model treats it as settled fact rather than a fringe claim.
Inside Claude's retrieval mechanics — and why they don't match ChatGPT's
ChatGPT is generally understood, across a wide range of public testing and vendor documentation, to lean more heavily on live browsing behavior as part of how it answers current or specific questions — Blyskal's dataset doesn't hand us a directly comparable invocation rate for ChatGPT, so we won't manufacture one here. What the Claude numbers do let us say cleanly is the contrast in kind: Claude's default posture is to answer from what it already knows and reach for the live web only when it judges that necessary, while ChatGPT's public behavior points toward reaching for the live web considerably more often as a matter of course. Same category of engine, structurally different retrieval habits.
| ENGINE | WHAT THE DATA SHOWS | WHAT IT IMPLIES FOR CONTENT |
|---|---|---|
| Claude | Live web search triggered on 36.6% of tested prompts; of citations pulled during a search, 79.2% traced to Brave's top 10 (Profound / Blyskal, Jul 22, 2026) | Most answers draw on training-time knowledge already baked in — being in the training distribution outweighs being freshly updated |
| ChatGPT | Understood to browse live considerably more often as a default behavior (Blyskal's dataset does not quantify a directly comparable rate) | A freshly updated page has a realistic shot at being fetched and cited inside the same cycle it was published |
A worked example: the same SaaS comparison page, two engines, two outcomes
This example is illustrative — a composite built to show the mechanics, not one client's actual data. Picture a B2B SaaS comparison page: "[Your Tool] vs. Three Competitors for Distributed Teams." It's been live for eighteen months, picked up backlinks from two review sites, gets referenced periodically in Reddit threads about tool selection, and this morning got a real update — new pricing, a revised feature table, current screenshots. That's a substantive refresh, the kind our earlier work on refreshing content for AI citations argues you should be doing on a cadence rather than as a one-off cleanup. The question this example is built to answer: does that refresh reach Claude and ChatGPT the same way?
Worth separating out the slice where Claude does search, because it isn't a black box either: 79.2% of the citations it pulls in that mode trace back to Brave's top 10. That's the mechanism where classic ranking discipline still applies to Claude directly — the work behind why comparison pages win a disproportionate share of AI citations still matters for that 36.6% slice of Claude traffic. But it's a minority slice. For the majority of Claude answers, the page's Brave ranking this week is beside the point, because Claude never went looking.
“The version of your page that matters to Claude, most of the time, is the one already sitting in its training data — not the one you shipped this morning.”
The strategic takeaway from the worked example isn't that the refresh was wasted. The refresh still matters for the buyers who read the page directly, for ChatGPT's more retrieval-heavy behavior, and for the 36.6% of Claude prompts where a search does fire and Brave ranking is in play. The takeaway is that the refresh was not the highest-leverage move available for winning Claude's majority case. The higher-leverage move for that case happened months earlier: getting the page's core claims corroborated widely enough, on enough independent sources, that they became part of what a model trained on the open web would already know.
What this means for your AI content strategy calendar
Most content calendars treat "refresh this page" as one action that serves every engine equally. Blyskal's numbers say that assumption breaks specifically at Claude. An AI content strategy built for 2026 needs two separate tracks running against the same page inventory, not one universal refresh rule applied everywhere. Something Inc.'s content marketing work increasingly builds those tracks in parallel rather than treating GEO as an add-on to the existing editorial calendar.
Freshness still pays — just not the way most content calendars assume
None of this contradicts the case for content freshness SEO discipline broadly. It narrows it. Freshness is a real lever for a real share of AI answers — for ChatGPT's more retrieval-heavy behavior, and for the minority of Claude prompts where a live search does fire. What Blyskal's numbers rule out is the assumption that freshness is a universal lever that pays off the same way across every engine your buyers use. Generative engine optimization stopped being one playbook the moment engines started diverging this visibly in how they retrieve, and Claude is the clearest single case of that divergence documented so far this year.
Do this next: pull your top ten citation-worthy pages, and for each one ask two separate questions instead of one. First, is this page fetchable and current enough to win a ChatGPT-style live fetch this week? Second, is the claim on this page corroborated widely enough, on enough independent sources, that it's plausibly already part of what a model like Claude learned in training, rather than something only your site says? Pages that pass the first test and fail the second are exactly the ones your calendar is currently over-serving. Pages that fail the first and pass the second are the ones it's ignoring. Fix the mismatch before the next refresh cycle, not after.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.