Keywords Outlived the Click

I The claim

Keyword research was built to win a visit: you found the phrase, ranked for it and counted who showed up. The visit is the part that broke. On a growing share of searches it never happens, and when it does your analytics may file it under the wrong name.

We still pick keywords for every offer that passes its test and goes on to be built. What changed is what we think the work is for, which is to tell us what people want and what they call it. We’d argue that was always the more valuable half, and that the traffic was a side effect we mistook for the product.

II Where the click went

Saharsh Agarwal and Ananya Sen, of the Indian School of Business and Carnegie Mellon, ran a randomised field experiment on 1,065 US desktop Chrome users, in which a browser extension hid Google’s AI Overviews for one group and left them on for another. Search Engine Journal reported the revised draft in July: organic clicks fell 39.8% on searches where the summary was shown.

It is a working paper on SSRN that has not been through peer review, and its participants were desktop users recruited from Prolific. We read it as a careful estimate from a narrow sample. It is also the only randomised one we have found.

The same paper checked whether the lost clicks were junk and found that they were ordinary visits from ordinary readers: quick returns to the results page and visits under ten seconds ran about the same in both groups.

For anyone still reporting on rank, this means the ranking report can stay green while the traffic falls, because position one now sits under a summary that has already answered the question.

III The dashboard is guessing too

Google rolled out AI reports in Search Console worldwide at the end of August. Per Search Engine Journal, they show impressions from AI Overviews and AI Mode by page, country and date, and at the time of that report they showed no click data. So you can learn that you appeared without learning whether anyone came.

ChatGPT is worse. LLM Pulse, a vendor that sells AI visibility tracking, reported that daily ChatGPT sessions across the sites it monitors fell about 46% between 9 and 11 October, while total traffic on those sites moved about 2%. It does not say how many sites that is. Its explanation rests on a third party’s analysis and OpenAI has not confirmed it: links in the new answers arrive without a tag or a referrer, so analytics files them as Direct. If that is right, the visitors are still arriving and have only lost their label.

We have stopped treating the referral report as ground truth, since channel attribution is an estimate with an error nobody has measured. Watch totals and what happens on the page you control, and ask new customers how they found you. Asking is a crude instrument. So is the dashboard, which won’t admit it.

IV Pick by intent

The Agarwal and Sen experiment split results by query type, and this is the part we act on. AI Overviews were triggered, or set to trigger, on 53% of informational queries, against 15% of navigational queries and 6% of transactional ones. The click loss sat in the informational group. Navigational and transactional queries showed no measurable change, though those samples were smaller.

So we sort the list by intent before we sort it by volume.

Transactional terms come from someone ready to buy, book or compare prices. Those still earn a page, and in our order of work that page already exists: the landing page is tested before anything is built, against a pass/fail number agreed in writing on the first day, and keywords are chosen afterwards for an offer that passed.

Informational terms come from someone trying to understand. We still want to know them, because they show what the market is confused about, but we no longer budget a visit against them. We write the answer plainly, with our name inside the sentence most likely to be lifted, and accept that being quoted is the return. Anyone forecasting pipeline from a how-to article in 2026 is forecasting from 2019.

V The words themselves

The second job of a keyword list is vocabulary, and the search trade is its own best example. 99Visibility pulled Google Ads volume estimates for the US, averaged over the twelve months to September 2026. These are bucketed advertiser estimates, so read them loosely. “AI SEO” averaged 6,600 searches a month, down 13% on the year before. “Generative engine optimization” averaged 4,400, up 35%. “Answer engine optimization” averaged 2,400, up 92%, and reached 4,400 in September.

The people who name things for a living cannot settle on a name for their own job, and a page written a year ago around the leading term is now written around the shrinking one.

This use of the research survives whatever happens to the click. The buyer’s word goes in the offer, the page headline, the founder’s posts and the release. In the Growth Unhinged survey of 300 software go-to-market leaders, 51% called SEO a core channel and 39% said the same of AI search, up 11 points in a year. We think both run on the same input, which is whether your words match theirs.

VI The one press release

Our method allows one press release, in the same later stage as the keywords. It has two readers and neither of them is the customer.

The first is a journalist. Cision’s 2026 State of the Media report, a vendor survey of nearly 2,000 journalists, found 66% rely on PR-provided content for story ideas. In the same survey 72% said fewer than a quarter of the pitches they get are relevant. So they do read releases, and most of what they read is shit.

The second reader is an answer engine, and it mostly reads the journalist. Muck Rack’s “What Is AI Reading?” study, as reported by PR Daily, put press releases at 1% of AI citations and journalism at 22%. Muck Rack sells PR software and the PR Daily piece does not give the method. Meltwater, another vendor reading its own tracking data, put releases at 0.2% of citations in May. Everything-PR followed 250 wire releases across five engines and found 47 cited by at least one within 90 days. Where engines did cite something, it was usually the coverage and rarely the wire URL. That publisher shares ownership with a PR firm, and its stated 15,000 observations do not match a design that multiplies out to 3,750. That makes three interested sources with soft methods, all pointing the same way, and the direction is all we take from them.

So we write the release as raw material. The claim is stated so that it can be checked and the numbers so that they can be lifted whole, the buyer’s vocabulary from the keyword list is in the first two sentences, and there are no adjectives a reporter would have to delete.

We don’t expect anyone to cite the release itself. It is there so that the people who do get cited have the right facts and the right words in front of them.