Life After the Search Box: Optimising for an AI-Mode-Default Google

TL;DR — On 19 May 2026 Google made AI Mode the default and rebuilt the search box for the first time in 25 years. The traffic consequences have been analysed to death. The consequence nobody has priced is quieter: the box was never only an input device. It was the mechanism by which the market told you what it wanted, in its own words, and Google published that back for free — because a keyword auction needed the data to exist. AI answers are not sold by keyword, so nothing in the new model produces that data as exhaust. Demand does not fall. It stops being observable.

Three layers of demand now exist: the queries you can still see, the queries people type that are withheld as too rare to name, and the synthetic sub-queries the engine writes for itself. The third decides whether you are cited — and no human ever typed it. You cannot target a query you cannot see, so 2027 planning shifts from choosing terms to covering question families, and from ranking against competitors to appearing across the sources a fan-out will retrieve.

What Google Actually Retired on 19 May 2026

At I/O on 19 May 2026, Liz Reid announced that AI Mode would become the default search experience globally, running on Gemini 3.5 Flash, and that the search input had been rebuilt for the first time in over 25 years. The new box expands as you type, accepts images, files, video and open Chrome tabs, and replaces autocomplete with AI-generated intent suggestions. Google reported AI Mode had passed one billion monthly users, with Personal Intelligence extending to nearly 200 countries and 98 languages.

The coverage oversold it, though. Google did not delete the ten blue links; they still render below and beside the generated answer. Nor is the surface yet where the volume is: Similarweb clickstream data covering January to April 2026 put AI Mode at 0.34% of all Google searches, measured before the default switched. The honest description of 19 May is not that search ended. It is that the input changed before the output finished changing, and the input is where your planning data comes from.

Google published its own behavioural report the same day. Shivani Mohan, VP of Data Science and UXR for Search, reported that the average AI Mode query is three times the length of a traditional query, that US follow-up queries were growing by more than 40% a month, and that planning-intent queries were growing 80% faster than AI Mode queries overall. Semrush clickstream analysis put numbers on the same shift: 7.22 words against 4.0 for traditional Google. Sundar Pichai framed search as becoming less a sequence of individual queries and more an ongoing conversation.

Those figures have been read as a content instruction — write longer, write conversationally, answer follow-ups. Not wrong, but downstream. They describe something that happens before any content decision: the unit of demand you have planned against for twenty years stopped being the unit the user produces.

The Consensus Playbook, Stated at Full Strength

The field converged on an answer fast, and it is a reasonable one: swap keyword research for prompt research, write in the language people actually use, structure pages so passages extract cleanly, add question-shaped headings, and track visibility in one of the new AI monitoring platforms. Same discipline, new interface.

That playbook has real evidence behind it and deserves stating at full strength before it is criticised. Prompt datasets now exist at scale: Semrush built a prompt database from clickstream covering more than 90 million real prompts; Profound computes visibility from a panel of over 400 million real user conversations. The prompt-keyword gap is closing, not widening — Semrush found the share of ChatGPT prompts phrased in traditional search language nearly doubled between October 2025 and February 2026, from 18.9% to 34.9%. Google shipped dedicated generative-AI performance reports in Search Console on 3 June 2026. And in the UK a regulator now compels disclosure: the Competition and Markets Authority imposed its publisher conduct requirement on the same day, having designated Google with strategic market status in general search in October 2025.

So the optimistic case is not naive. Tooling is arriving, regulation is arriving, and user language is drifting back toward something a keyword tool recognises. The problem is that every one of those developments measures the same thing — your presence — and none returns what was lost, which is the question.

Read the Search Console release carefully. The generative-AI reports give impressions, pages, countries, devices and dates: no clicks, no click-through rate, no average position, no query dimension. Google was explicit that the data already sat inside the overall performance report; what shipped was a view, not a new measurement. You can see that you appeared. You cannot see what you appeared for.

The Box Was a Disclosure Mechanism, Not an Input Field

For twenty-five years the search box performed two functions nobody bothered to separate, because one artifact served both. It was an input device: a place to state what you wanted. It was also a disclosure device — the only place in commercial life where millions of buyers stated their intent in their own words, in a form that was counted, aggregated and published back to the people selling to them.

The second function built the profession. Keyword volume, demand curves, seasonality, gap analysis, the whole apparatus of competitor keyword and backlink analysis — all of it is downstream of a single fact: that the string a person typed was recorded, counted, and made available. Remove the disclosure and the demand does not disappear. Your ability to observe it does.

Why the demand data existed at all

Google did not publish query volume as a courtesy to marketers. Keyword volume is documentation for an auction. Advertisers cannot bid on what they cannot see, forecast spend against a term with no volume estimate, or commit budget to inventory they cannot size. Search volume was public because the ad product was sold by keyword, and a keyword marketplace requires published quantities. Organic planning free-rode on that requirement for two decades.

Aaron Wittersheim and Tom Lustina of Straight North put it sharply in March 2026: query data will be released when advertisers are paying for it, and not before. That is not cynicism. It is the causal mechanism behind every keyword tool you have ever used.

Why nothing in the new model reproduces it

Now apply that forward. Sponsored placements inside AI answers are matched contextually, against the substance of a conversation, not bid against a keyword string. An engine selling attention inside a generated answer needs no public vocabulary of biddable terms, because there are no terms — there is a conversation and a model deciding what is relevant to it. The pressure that forced query volume into the open has no equivalent in the new stack.

So the correct expectation for 2027 is not that query data is late. It is that query data is structurally absent. Nothing in the AI answer business model generates it as exhaust, so nothing will publish it. Any demand data you get from here is bought from a panel, inferred from a model, or collected by you.

The Three Layers of Demand in 2027

Separate the two functions of the box and demand splits into three layers with different authors, different visibility and — this is the part that matters — different jobs. Read the last column first.

Layer two was already half-dark before AI Mode existed

The most important corrective here is that the visibility problem predates the interface change. Ahrefs analysed 22 billion clicks across 887,534 properties and found 46.77% of clicks arrive on queries Search Console will not name, with most individual sites between 45% and 80%. That used April 2025 data — before AI Overviews and AI Mode scaled. The 2022 figure was 46.08%. Half your demand has been dark for years.

Google’s rule withholds queries not issued by more than a few dozen users over a two-to-three month window, plus anything containing personal or sensitive detail. Kevin Indig measured it from the other side in February 2026, comparing Search Console’s aggregate endpoint against its query-dimension endpoint across roughly 450 million impressions on ten B2B SaaS properties: about 75% of impressions and 38% of clicks filtered out, with a site-level range of 59.3% to 93.6%.

Notice why an AI-default interface worsens this automatically rather than deliberately. The threshold is a function of rarity. Longer queries are rarer; more personal queries are rarer still and likelier to trip the sensitivity filter. Google has told us AI Mode queries are three times longer and increasingly conversational, and that Personal Intelligence now connects Gmail and Photos to the search context in nearly 200 countries. Each of those pushes more of your demand below the naming threshold. Nobody decided to hide it. The privacy rule did what it was built to do, applied to a population of queries that shifted underneath it.

Layer three is the one that decides citation

The synthetic layer differs in kind, not degree. When a prompt arrives, the engine decomposes it and retrieves against the parts. Seer Interactive and Nectiv found an average of nine to eleven sub-queries per prompt, with 59% triggering between five and eleven and roughly a quarter triggering twelve to nineteen — occasionally twenty-eight.

Those sub-queries are what your page is judged against. They are written by a model, they vary between runs, and they exist only inside a retrieval pipeline. No privacy rule hides them and no regulator has asked for them. They were never human demand, so they can never appear in a dataset of human demand — at any price, under any regime. That is the hardest fact in this article, and the one that decides what optimisation can now mean.

Key takeaway. Your reporting stack measures layer one. Your enquiries increasingly come from layer two. Your citations are decided in layer three. One dashboard, pointed at the smallest and least commercially important of the three.

Measuring What You Have Lost: The Dark-Demand Ratio

Averages are useless for planning because the variance between sites is enormous. Your own number takes five minutes. Search Console will not tell you what the hidden queries say, but it will tell you how much it is hiding, because the arithmetic gives it away: chart totals include every click, and the query table excludes anonymised ones. The gap between them is the measurement.

THE DARK-DEMAND RATIO

Open Performance in Search Console, set a three-month window, note total clicks, then export the queries table for the same window and sum its clicks. The ratio is (total clicks minus summed query clicks) divided by total clicks. Run it property-wide, then filtered to each of your three highest-value page groups.

Below 30%: your demand is still largely legible. Keyword-led planning still works; keep it, and watch the trend.

30% to 55%: the mainstream case. Plan by question family rather than by term, and stop ranking content ideas by volume you can only observe for part of the market.

Above 55%: you are already operating blind. Optimising the visible tail means optimising a sample you did not choose. Reconstruction is the priority, not refinement.

Then run it again for the same quarter a year earlier. The level tells you your situation; the slope tells you how fast your planning inputs are decaying. A site whose ratio moved from 34% to 58% in a year has not lost traffic — it has lost the ability to explain the traffic it still has, which is a different emergency and takes longer to fix.

A rising ratio is not automatically bad news — it signals demand becoming more specific, and specific demand converts. It is unambiguously bad for planning. You cannot brief a writer, forecast a quarter or defend a budget against a query set describing less than half of your traffic.

How much of my Search Console query data is actually missing?

Just under half of clicks arrive on queries Search Console will not name — 46.77% across 22 billion clicks in Ahrefs’ analysis, with most sites falling between 45% and 80%. Impressions are hidden at a far higher rate. The traffic is counted in your totals; only the query text is withheld. Compute your own figure, because the spread between sites is wider than the average is informative.

Rebuilding a Demand Signal You Own

If the public utility is switching off, the replacement has to be private. Four substitute sources are available to almost any business, and the useful way to see them is not as tactics but as four biased instruments pointed at the same object. Each distorts in a known direction, which means each can be corrected — and the corrections are the method.

Your site search log is the cleanest source of customer vocabulary you will ever own and the most misleading source of volume. Everyone in it has already found you, so it under-represents the questions people ask before they know you exist — the category-entry questions where a generated answer decides whether you make the shortlist. Use it for phrasing, weight it by new versus returning visitors, and never treat its frequencies as market demand.

Your sales and support conversations carry questions in the exact form buyers ask them, including constraints and objections no keyword tool has recorded. They over-represent late-stage intent and, more subtly, the questions your own people prompt. The correction is to mine the losses: enquiries that went cold, calls that never came. Those are the questions an engine answered on your behalf — the most valuable and the hardest to collect.

Community and forum mining — trade forums, subreddits, industry groups, the comment sections under listicle placements in your category — gives you constraint language and edge cases, because people post when something has gone wrong or an option has been ruled out. It over-represents the unusually engaged and the aggrieved. Use it to discover question shapes, never for frequency. Sources like Hacker News and similar high-signal communities are especially useful here because the objections are stated explicitly rather than implied.

Finally, AI visibility platforms — Profound, Scrunch, Peec, Semrush’s AI toolkit and the rest — belong in the panel, provided you understand what they sample. Microsoft’s Clarity team put it precisely in March 2026: synthetic prompt simulation measures what a model could say when asked, not what it did say when a real person asked, and two identical prompts minutes apart can return different citations. Brainlabs was blunter in April 2026: prompt volume figures are probabilistic estimates, and the number you most want — how many people saw an answer mentioning you — is unknowable. Use these tools to benchmark presence over time, never as a volume source, and check whose panel you are buying: a UK specialist reading a US consumer panel is reading someone else’s market.

What none of them gives you — and why that is the right answer

Not one of the four returns a volume figure. That feels like failure until you ask what volume was for. It answered a prioritisation question: of the hundred things I could write about, which first? That assumed a finite list of targetable units competing for a finite budget.

In a fan-out world the question changes shape. The engine decomposes the prompt whether or not you optimised for the decomposition, and retrieves against every part. So the operative question is not which question to answer first. It is whether there is a question in this family your estate cannot answer at all — because an unanswerable sub-query is a hole through which a competitor gets retrieved. Coverage replaces prioritisation, and coverage needs no volume estimate.

Key takeaway. Four biased instruments, honestly corrected, beat one precise instrument pointed at the wrong layer. Site search gives vocabulary, lost deals give the eliminating questions, communities give constraints, visibility tools give trend. None gives volume — and volume answered a question you no longer have.

From Targeting to Coverage: What This Means for Links

Everything above is diagnosis. It lands awkwardly for a discipline that spent two decades getting more precise: when you cannot aim, the only remaining strategy is to cover. Two kinds of coverage matter, and they are bought differently.

Coverage of the decomposition

The first is on your own estate, and it is a consolidation move rather than an expansion one. A page built to rank for one term answers one question well and its neighbours badly. A page built to serve a question family answers the sub-queries a decomposition is likely to generate — price, compatibility, timeline, what-happens-if-it-goes-wrong — with explicit, extractable claims rather than allusions. Which is why thin pages targeting close variants now underperform a single thorough one: the variants were never the unit. They were an artefact of a matching system that no longer does the matching.

The test is unglamorous. Take a question family, write out every sub-question a reasonable decomposition would produce, and check which your estate answers with a specific claim rather than a gesture. The gaps are your plan. It is the same instinct behind building genuinely useful interactive assets — a calculator answers a whole family of quantified sub-questions in one artifact, which is exactly what a retrieval pipeline is looking for.

Coverage of the retrieval set

The second kind cannot be bought on your own domain at all, and this is where the argument becomes a link-building one rather than a content one. Each sub-query runs its own retrieval, your page is one candidate among many, and organic position is a weak predictor of whether it wins. Surfer’s analysis of 173,902 URLs across 10,000 keywords found 68% of pages cited in AI Overviews were not in the top ten organic results. Ahrefs compared 730,000 query pairs and found AI Mode and AI Overviews reach semantically similar conclusions 86% of the time while citing the same URLs only 13.7% of the time — two surfaces from one company, same answer, almost entirely different sources.

That is a portfolio problem with a portfolio answer. If the query distribution is unobservable and the retrieval outcome is high-variance even at fixed quality, the rational response is to hold positions across many independent sources rather than concentrate on your own. Presence on the trade title, the association resource page, the comparison roundup, the practitioner directory and the community thread is not vanity link acquisition; it is coverage of retrieval paths you cannot enumerate. The classical case for earned links used to be authority transfer. The 2027 case is variance reduction against a demand distribution you are no longer permitted to see.

A second, independent route reaches the same conclusion, and it is the one to take to a finance director. Look again at which demand stays visible. Adido’s analysis of UK client data found that while the impression-weighted midpoint of query length rose about 13% year on year, the click-weighted midpoint fell — from 3.75 words to 3.0. The queries still producing clicks are getting shorter, not longer, which in practice means branded and navigational. Semrush found the same pattern from the prompt side: among prompts that do match traditional search language, most carry navigational or transactional intent.

So the last reliably legible demand signal you have is people typing your name. That is not a consolation prize but a strategic instruction. Branded search is the one thing that stays in layer one, and it is manufactured almost entirely off your own site — by being mentioned, cited, recommended and argued about in places you do not own. The measurement collapse and the link building strategy answer point in the same direction, which is usually a sign the analysis is sound rather than convenient. Practitioners tracking how AI Overviews interact with backlinks have been converging on this from the citation side for a year.

Where This Argument Could Be Wrong

The strongest counter is not that demand is fine. It is that a reporting lag is being dressed up as a structural loss, and it arrives with four pieces of evidence.

First, scale: AI Mode was 0.34% of Google searches in the January-to-April 2026 window, so most demand is still typed into a box and still reported. Second, convergence: prompts phrased like search queries nearly doubled as a share in four months. Third, panels: Profound and others do observe real prompts at scale from opted-in participants, which is not simulation. Fourth, regulation: the CMA’s publisher conduct requirement compels controls, attribution with clear links and detailed engagement metrics, phased through December 2026 with page-level grounding controls to March 2027. On that reading, the data is coming.

Each is true, and the first two genuinely bound the claim. The 0.34% figure deserves particular respect: it predates the default switch, but it is a reminder that behaviour changes far more slowly than interfaces do, and anyone forecasting a cliff is guessing. If you sell to a market that still types three words into a box, most of this is a 2028 problem.

But the bounds do not reach the core claim, for four reasons. The anonymisation gap was already 46.77% in April 2025, before AI Overviews scaled and while nearly all demand was still box-shaped — so the loss of query visibility is not caused by AI Mode and will not be reversed by AI Mode staying small. Convergence is happening at the entry turn, while follow-ups grow more than 40% a month, and the fan-out layer is untouched by any convergence in human phrasing. Panels observe the user’s prompt, never the sub-queries generated from it, and panel composition is rarely your market. And the regulatory reading is the most instructive of all: read the conduct requirement and you find controls, attribution and engagement metrics. Nowhere does it require query data. The world’s first binding remedy on this treats the missing thing as credit, not knowledge. Nobody is asking for the questions back, because nobody has noticed they were the asset.

Two things would falsify this argument. The first is any engine shipping a fan-out dimension — a report naming the synthetic sub-queries a URL was retrieved for. That would make the decisive layer targetable and most of this obsolete; the CMA’s December 2026 and March 2027 milestones are where to look for anything query-shaped. The second is replication evidence that fan-out is near-deterministic. If one prompt reliably produces the same sub-queries across runs, accounts and regions, layer three collapses back into layer one and can be reverse-engineered. Current evidence runs the other way, but that is an empirical question and data will settle it.

Can you still do keyword research for AI Mode?

Yes, but for a different purpose. Keyword data still tells you how a market describes its problems and which vocabulary carries commercial intent, and it is still the fastest way to map a topic. What it no longer does is tell you what to target, because the retrieval decision is made against sub-queries no keyword tool contains. Treat it as a language input to a coverage plan, not as the plan.

Worked Example: Stourgate Hearing

Stourgate Hearing is an independent audiology group headquartered in Shrewsbury, with 19 clinics across Shropshire, Cheshire and the Welsh border, £11.4m of revenue and 34 staff. Its 240-page site included a 90-page symptom-and-support library producing roughly 60% of enquiries. The category stress-tests everything above, because hearing loss generates exactly the long, personal, slightly embarrassing question a privacy threshold is designed to suppress.

In Q4 2025 the property took 9,400 organic clicks and the query table named 6,100 — a dark-demand ratio of 35%. In February 2026 the marketing lead built the FY27 plan on 340 tracked keywords. Clicks were down 18% year on year while 61% of those keywords had held or improved position. The board asked the obvious question and the report could not answer it.

In March 2026 they ran the ratio by page group rather than property-wide, which is where it became useful. Property-wide it had risen to 52%. On the symptom library it was 74%; on brand and product pages, 21%. The dark demand was concentrated precisely on the pages that generated enquiries, so the reporting was clearest exactly where the business was weakest.

Reconstruction ran from March to June across four sources: 11,000 site-search queries a quarter, 900 consented call recordings transcribed from April, a question log kept by audiologists during appointments, and community mining across a UK hearing-loss forum and two subreddits. The output was 71 recurring question families; the library answered 23. The other 48 were almost all constraint questions rather than topic questions — whether a device works with spectacles, survives swimming, is covered by an employer’s scheme, or whether an NHS assessment delays a private fitting. None appeared in the 340-keyword set, and 6 of the 71 appeared in any keyword tool at all.

Between April and July they consolidated the 90 thin pages into 31 family pages, each carrying explicit numbers — price bands, fitting timelines, device compatibilities, warranty terms — rather than referring readers to a consultation. In parallel they ran an earned-placement programme aimed at the sources a hearing decomposition would plausibly retrieve: two audiology trade titles, a regional newspaper health column, a national hearing-loss charity, three clinician directories and their local NHS trust’s patient-information page.

Measured in late July 2026: organic clicks 7,100, still 24% below the Q4 2025 baseline. The generative-AI report, available in the UK beta from 18 May data, showed 14 of the 31 family pages appearing in AI features. Enquiry completions were up 12%, phone enquiries 19%, branded search 38%, referring domains up 26 in five months. The traffic model did not recover and they stopped forecasting that it would.

The honest negatives are instructive. The panel over-weighted people who had already chosen them, describing the questions of customers rather than of the people an answer engine had already eliminated; lost-enquiry callbacks were the only corrective and produced the thinnest data of the four. Two sources took five months to yield usable volume because the call-transcription consent workflow had to clear a data-protection review. Consolidating 90 pages into 31 cost positions on more than 40 long-tail terms that were still converting — accepted knowingly, and it still hurt in Q3. And four months went on a prompt-tracking platform whose panel was almost entirely American, producing a share-of-voice figure that moved independently of anything they did.

What To Do On Monday

Eight things, in order, none needing budget approval.

• 1. Compute your dark-demand ratio property-wide and for your three highest-value page groups. Write each number down with the date — you are starting a time series.
• 2. Recompute all four for the same quarter last year. The slope is the finding, not the level.
• 3. Open the generative-AI performance report if your property has it and list the pages appearing in AI features. Then write down what it does not tell you, so nobody mistakes an impression counter for a demand signal.
• 4. Take one page group and list the question families it serves rather than the keywords it targets — constraint questions, not topic questions.
• 5. Pull last quarter’s site-search log and the last 50 enquiry calls, and extract every question in the customer’s own words. Do not tidy them into keywords; the untidiness is the data.
• 6. Mark the questions on that list your estate cannot answer anywhere. That is your content plan for the quarter, no volume estimates required.
• 7. Name the five third-party sources most likely to be retrieved for your category’s sub-questions and check whether you appear on any. Where you do not, that is a link building brief, not a content brief.
• 8. Retire one report that treats a query as a targetable unit; replace it with a coverage report — families served, families with gaps, third-party sources held. Track link velocity against that map rather than against a keyword list.

The profession spent twenty-five years getting better at aiming, and the target has stopped being visible. That sounds like a catastrophe and is mostly a reallocation. Precision was always borrowed — it existed because an ad auction needed published quantities, and it was withdrawn the moment the auction stopped being priced in keywords. What replaces it is older, duller and much harder to fake: know your customers’ questions because you collected them yourself, answer the whole family rather than the term, and be present in enough places that you get retrieved by paths you never mapped. You cannot aim at a question nobody typed. You can be standing where the answer gets assembled.

Meta title: Google AI Mode Optimisation: Life After the Search Box  |  Meta description: Google made AI Mode the default in May 2026. The real change is that demand stopped being observable. How to measure what you lost and plan coverage instead of keywords.