SEO did not die, its unit of work changed
Search engine optimisation still works in 2026. What changed is what you are optimising: not a page that has to rank, but a passage an answer engine can lift and attribute. AI Overviews now appear on roughly 48% of Google queries, up from about 34.5% in December 2025, so on half of all searches your page is competing to be quoted before it competes to be clicked.
The click data tells the rest of the story. Seer Interactive, tracking 53 brands, 5.47 million queries and 2.43 billion impressions from January 2025 to February 2026, measured organic click-through of 0.61% on queries carrying an AI Overview against 1.62% on queries without one. Ahrefs, across 300,000 keywords, found the top-ranking page loses roughly 58% of its CTR when an Overview sits above it. Pew’s panel study put it in human terms: 8% of users click an organic result when an AI Overview is present, versus 15% when it is not.
The strongest evidence is the one built as an experiment rather than an observation. A randomized field experiment run by researchers at the Indian School of Business and Carnegie Mellon measured a 38% drop in organic clicks on affected queries, with the zero-click rate rising from 54% to 72%.
Read those four studies together and the conclusion is narrow, not apocalyptic: the same visibility now yields fewer visits. Your work is not less valuable. It is less visible in the one report you used to open first.
v1be is a company that builds GEO and SEO focused AI growth software; our flagship engine, Conty, researches the live web, writes long-form articles and carries its own citations. This piece is the operating manual we use for the transition, and none of it requires a tool you cannot buy or build this quarter.
Read the CTR studies properly before you restructure the team
The traffic loss is real but it is not evenly distributed, and treating it as a flat tax across the site leads to the wrong cuts. The compression concentrates on queries an AI answer can close in one paragraph: definitions, quick how-tos, unit conversions, symptom lists, “what is” and “how long does”. Those were often your highest-volume, lowest-value pages.
Three distinctions worth making before anyone rewrites a strategy deck:
- Informational versus transactional. A query that ends in a decision (pricing, comparison, “near me”, “book”) still sends the click, because the answer engine cannot complete the transaction inside the answer.
- Unbranded versus branded. Branded queries barely move. If demand for your name grows, the AI layer routes it rather than absorbing it.
- Volume versus value. Losing 40% of clicks on definitional content while holding conversions flat is a reporting problem, not a revenue problem. Check the conversion line before you panic about the sessions line.
There is also a countertrend worth knowing so you do not over-correct: the decline through 2025 did not continue in a straight line, and CTR on AI Overview queries recovered measurably in early 2026 as the interfaces matured and started sending more traffic on complex questions. Plan for compression, not for extinction.
Your instrument panel was replaced mid-flight
The report changed before the job did. Impressions without clicks is the new normal reading.
On 3 June 2026, Google added Generative AI performance reports to Search Console, rolled out UK-first under a binding order from the UK’s competition regulator. The report shows impressions inside AI Overviews, AI Mode and AI features in Discover, broken down by page, country, device and date. Data starts 18 May 2026, with no historical backfill.
Note what is missing in version one: no clicks, no CTR, no average position, no query-level breakdown. You can see that you appeared. You cannot see what appearing did.
That constraint dictates how to use it. Treat AI impressions as shelf space, not as traffic, and pair them with a demand signal that does convert:
- Baseline the impressions per page, then watch which pages hold their appearance rate after each content update. Movement here is your citation-worthiness signal.
- Read it next to branded search and direct traffic. If AI impressions climb while branded queries and direct sessions climb with them, the answer layer is doing the work of an ad you did not buy.
- Stop reporting sessions as the headline number. A dashboard where organic sessions sit alone at the top now tells the board a story that stopped being true in 2025.
The general version of this measurement problem, for engines that publish no report at all, is covered in our guide to measuring AI brand visibility. One market already went further than Google: Yandex publishes an actual Share of Voice metric, which we broke down in getting cited by Russian AI search.
Keyword research becomes question research
The keyword still matters as a demand signal, but it stopped being the shape of the work, because people do not type keywords into an answer engine. They type sentences, then follow up. Optimising for “crm pricing” and optimising for “which crm makes sense for a 12-person agency that already uses HubSpot forms” are two different jobs, and only the second one describes what your buyer actually asked.
What that changes in practice:
- Cluster by question, not by keyword. One page should close a primary question and the four or five follow-ups that reliably come after it. Those follow-ups are where the citation usually lands.
- Mine the long tail you cannot see. Conversational queries are too varied to show up individually in volume tools, so read your own sales calls, support tickets and site search logs. That is first-party demand data your competitors do not have.
- Map the secondary intent. A query about a mechanism (“how does X work”) almost always carries a hidden decision intent (“should I use it”). Pages that answer only the first half get quoted; pages that answer both get quoted and clicked.
Write claims, not paragraphs
The extractable unit is a two to four sentence block that answers one question and stays true when it is cut out of the page. That is the single highest-leverage writing change of the last two years, and most teams still have not made it.
Four properties make a passage liftable:
- It repeats its subject. “The platform starts at $149 per month” survives extraction; “It starts at $149” does not, because the pronoun loses its antecedent the moment the sentence moves.
- It carries a checkable token. A number, a date, a named source. Models repeat what they can corroborate elsewhere.
- It sits directly under the heading it answers. The first paragraph after an H2 is the highest-value real estate on the page.
- It says something the other results do not. Information gain is the one ranking concept that got more important, not less. Reproducing the current top result gives the model a second copy of an answer it already has.
Freshness stopped being cosmetic too. A page carrying 2024 pricing does not merely look stale; it teaches the model a wrong fact about your company and keeps teaching it until the page changes.
The crawl economy: decide who you let in
Crawling and referral used to be roughly balanced. They are not any more, and the ratio is now a business decision rather than a technical footnote. Cloudflare Radar data through 2026 shows Googlebot crawling around five times for every visitor it sends, PerplexityBot in the hundreds, GPTBot around nine hundred, and ClaudeBot in the thousands. In May 2026, Cloudflare attributed 51.8% of AI crawler requests to training and only 9.3% to search.
Two honest conclusions follow, and they point in different directions:
- If you sell attention, publishing is a business where that ratio is a straightforward transfer from you to a model vendor, and restricting training crawlers while allowing search crawlers is a defensible position.
- If you sell a product or service, the calculation inverts. Being absent from the answer means being absent from the shortlist, and no ad budget buys back a recommendation you were not part of. Stay open, and spend the energy on being the source worth quoting.
Either way, make the decision explicitly. The worst outcome is a robots.txt that blocks the crawlers feeding your buyers’ shortlist because someone copied a template.
What did not change (and why that is the useful part)
Most of the discipline survived intact, which is why teams that kept their fundamentals are having a much calmer 2026 than teams that chased tactics. Still load-bearing:
- Crawlability and rendering. Content that arrives only after JavaScript is content that may never be read, by either kind of engine.
- Intent match. Answer engines are ruthless about it, because they are optimised to close the question rather than to fill a page.
- Speed and structure. Clean heading hierarchies now serve extraction as well as accessibility.
- Links, as corroboration. They stopped being a scoreboard and became evidence: independent sources saying consistent things about the same entity is exactly what a model checks before repeating a claim.
- Brand demand. Branded search is the most AI-resistant asset you own, and the one line in your analytics that generative answers tend to raise rather than absorb.
The deeper continuity: every one of these was always about being the most useful, most verifiable answer. The machines simply removed the margin for pages that were none of those things but ranked anyway. The broader comparison sits in GEO vs SEO: what changes.
What the SEO week actually looks like now
The cadence changed more than the craft: measure appearances weekly, rewrite passages continuously.
The routine that produces results in 2026 mixes classic technical hygiene with a citation loop that did not exist three years ago:
- Monday: read appearances, not just sessions. Search Console AI impressions per page, branded search volume, direct traffic. Note which pages gained or lost their appearance rate.
- Tuesday: run the prompt panel. A fixed set of 20 to 40 buyer questions across the engines your market uses. Log which brands and sources come back. This is your share of answer, and it moves slowly enough that weekly readings are meaningful.
- Midweek: fix the passages. Take the pages that rank but are not quoted and rewrite the first paragraph under each heading into a self-contained answer with a checkable number in it.
- Thursday: publish information gain. One piece per cycle carrying something no one else has, whether that is your own data, a documented process or a tested result. This is the only durable moat left.
- Friday: entity maintenance. Keep your description, pricing and claims identical across your site, your platform profiles and your listings. Contradictions are the fastest way to be dropped from an answer.
Thirty days of this will not rebuild traffic that the interface absorbed. It will tell you which of the three constraints (ranking, extractability, corroboration) is actually binding on your domain, which is the diagnosis most teams are missing while they argue about whether SEO is dead.
The honest summary: the channel got harder to report and no less valuable to own. The brands that will look lucky in 2027 are the ones that spent 2026 becoming the answer rather than defending the ranking. Conty writes this class of content on a research, write, review pipeline, with the sources attached and a person approving before anything publishes.
