TL;DR
• In the first four months of 2026, 68.01% of US Google searches ended without a click to anywhere — no website, no ad, not even a Google property, against 60.45% in 2024.
• The industry’s answer is to re-base the click-through-rate term and keep everything else. That repairs a coefficient, not a specification.
• Three terms did not merely fall — they changed meaning. Volume stopped being observable, click-through became a selection filter rather than a thinning parameter, and conversion rate stopped being a property of your website.
• The last is the dangerous one, because it moves in the flattering direction. Adobe measured AI-referred visitors converting 38% worse in March 2025 and 42% better a year later. No website improved by eighty points in twelve months.
• THE TRAFFIC-MODEL TEARDOWN sorts every term into level-shifted, meaning-changed or discontinued. THE FOUR SIGNATURES reads sessions, rate and count together, so you can tell filtration from displacement before the revenue line does.
• The rebuilt model forecasts counts not rates, commits to supply instead of predicting demand, and treats earned corroboration as the only line item you can still put a date against.
1. The forecast that stopped forecasting
Every agency, in-house team and consultant has built the same object. Take a set of keywords, attach a monthly search volume to each, multiply by a click-through rate for the position you expect to reach, multiply again by a conversion rate and an average order value, and hand the result to somebody who controls a budget. Nobody funds a content programme on a strategy deck; they fund it on a number with a pound sign in front of it.
On 8 June 2026, SparkToro published the analysis everyone has been quoting sincezero-click headline number, built on Similarweb clickstream data covering US Google searches from January to April. It found 68.01% of searches ending without a click of any kind, against 60.45% in 2024 and around 45% in 2016. The share generating at least one click fell 9.51 points, a 22.9% decline; the share leading to another Google search instead rose 7.2 points. Roughly 276 of every 1,000 searches now send a person anywhere at all.
The caveats cut both ways: the long series stitches together three different panels, and SparkToro sells audience research to people who have concluded traffic is no longer the point. A second series is harder to wave away. Ahrefs tracks 75,000-plus opted-in domains, and the share of traffic Google sends them fell eight points between June 2025 and May 2026 — measured on sites with professional marketers actively trying to grow traffic, the population best equipped to resist it.
The repair the industry actually made
The response has been swift, competent and almost entirely concentrated in one term. Open any 2026 forecasting guide and you find a version of this: projected clicks equal search volume, times a click-through rate for the target position, times an AI-Overview adjustment, times a probability of reaching that position, discounted by a confidence margin. The working assumption this summer is a 40–60% click-through reduction where an AI Overview triggers, and tools now ship discount tables: AI Overview −58%, shopping pack −45%, featured snippet −35%.
The better practitioners go further, and rightly. Pull your own click-through rate from Search Console rather than a benchmark built on another brand’s recognition. Segment branded from non-branded, informational from transactional. And stop issuing point forecasts — publish scenarios with assumptions named, because a single number here is, as one 2026 guide puts it, professional malpractice.
None of that is wrong, and the miss it corrects is real: a May 2026 study of 4.5 million impressions found published benchmarks overstating non-branded clicks by five to twenty-four times.
2. Stale is not the same as misspecified
There are two ways a model can be wrong, and the difference between them decides whether recalibration fixes anything.
A stale model has the right structure and the wrong numbers. Its variables still mean what they used to; the coefficients have drifted. Re-estimate and the model is good again. Staleness announces itself: forecasts miss in a consistent direction by a consistent magnitude, and the miss shrinks the moment you update the inputs.
A misspecified model has the wrong structure. Its variables no longer measure what their names imply, or the relationships between them no longer hold. Re-estimating one does not repair it. It does something worse: it produces fresh, tightly fitted coefficients that make the model look calibrated while it continues to answer a question nobody asked.
You can re-estimate a coefficient. You cannot re-estimate a definition.
A stale model is wrong and looks wrong. A misspecified model is wrong and looks calibrated, and the second is more expensive because it survives the meeting.
What is a zero-click traffic model?
A zero-click traffic model is a forecast built for a market in which most demand resolves without a visit. It differs from a recalibrated conventional model structurally, not in its coefficients: it forecasts absolute counts of business outcomes rather than rates or sessions; it treats search volume and click-through as diagnostics rather than the base of the projection; and it puts committed inputs — assets published, sources earned — on the forecast line, expressing demand-side conversion as a band drawn from its own history.
3. The traffic-model teardown
Before rebuilding, take the model apart term by term. Three verdicts matter, and the second is the one people skip.
Level-shifted — the term still measures what its name says and the value moved. Re-estimate and carry on.
Meaning-changed — the arithmetic still runs, but the quantity is now a composite of your performance and somebody else’s product decision, inseparable from your own data. It cannot be held constant, forecast forward, or read as performance.
Discontinued — the term refers to something no longer observable at the coverage the model assumes. It can be printed. It cannot support a forecast.
| Term in the classic model | What it used to measure | What it measures in 2026 | Verdict and treatment |
| Monthly search volume | Total stated demand for a phrase | A shrinking, unevenly sampled slice of stated demand, estimated from vendor panels | DISCONTINUED as a base. Use for relative sizing between question families only. |
| Click-through rate by position | A thinning parameter — roughly the same coin flip for every user at that position | A selection filter set by how completely the answer above resolved the need | MEANING-CHANGED. Estimate per question family from your own data, never per position from a curve. |
| Sessions | Arrivals, a fair proxy for demand reached | The residue of demand the answer could not close | MEANING-CHANGED. Demote from headline KPI to diagnostic input. |
| Conversion rate | The persuasive power of your site | Your persuasion times the engine’s filtration, with no way to separate them | MEANING-CHANGED, and the most dangerous term here. Never hold constant, never celebrate in isolation. |
| Revenue per session | Value extracted per visit | Value per visit, inflated by whichever visits were removed upstream | MEANING-CHANGED. A rise is as consistent with shrinkage as with improvement. |
| Absolute conversions and revenue | The outcome | The outcome, unchanged | INTACT. The only clean quantity left. Promote to the top of the model. |
| Branded and direct demand | A separate channel; usually a fixed baseline | Where resolved-but-not-arrived demand lands after the answer names you | PROMOTED. Becomes a dependent variable of your visibility work, not a constant. |
| Third-party sources carrying an accurate statement about you | Absent, or modelled only as an input to rankings | Supply that unnamed retrieval steps draw on, and a production variable you control | NEW LINE ITEM. The only term in the model you can put a date against. |
Run this before you touch a coefficient. In most cases four rows come back amber, the two green rows at the bottom are not in the model at all, and the recalibration project everyone has scheduled touches exactly one row.
Key takeaway
Sort every term before you re-estimate anything. Only level-shifted terms are repaired by new numbers; meaning-changed terms get demoted to the diagnostics. The two that survive intact — absolute outcomes, and the corroboration you cause to exist — are usually the two your model does not contain.
4. The term nobody is fixing: your conversion rate stopped being yours
Here is the finding the industry has spent 2026 celebrating. Adobe Analytics reported that in March 2026 visitors arriving at US retail sites from AI assistants converted 42% better than non-AI traffic and generated 37% more revenue per visit. Twelve months earlier the same channel converted 38% worse — an eighty-point swing in a single year.
The received reading is that AI traffic is high-intent: buyers compare inside the chat and arrive ready. True as far as it goes. What it misses is what an eighty-point swing says about the parameter rather than the channel. No retailer’s checkout improved by eighty points. The population being passed through changed, because a third party altered how much of the need it resolved before letting anyone leave.
The spread across studies sharpens it. Published conversion advantages for the same year run from Visibility Labs at roughly 1.3 times non-brand organic across 94 ecommerce brands, through ALM Corp at 31% and Adobe at 42%, to Emarketed at 4.4 times and Opollo at 14.2% against 2.8% across 312 B2B firms. A parameter whose estimates range from 1.3x to 5x within one calendar year is not a property of anything you own.
What is actually doing the work
Click-through was a thinning parameter. Some fraction of the people who saw your result clicked, and that fraction was close enough to independent of who they were that treating it as a coin flip worked for a decade. That is why the multiplication chain held: each stage removed a random-ish share of the last, so a rate estimated on last year’s population was fair for next year’s.
An answer engine does not thin. It selects, and on precisely the variable your funnel cares about. It resolves what it can — the definition, the comparison, the how-does-this-work — and passes through what it cannot: buy this, book this, check availability, decide whether to trust this supplier. Seer Interactive’s 2026 study across 53 brands and 2.43 billion impressions found AI Overviews on around 36% of informational queries and 5% of transactional ones. The filter is not uniform; it is aimed. Remove the low-intent arrivals and the conversion rate of what remains must rise, whatever anyone does to the website — and the filtered visits never appear in your analytics at all.
One 2026 arXiv paper tried to separate the two properly and found a raw association near 5.7 times collapsing to a net causal estimate around 1.82 times, 95% confidence interval 1.31 to 2.54. There is a real effect. It is about a third of the headline. Any model booking the headline as its conversion assumption over-forecasts that term threefold.
Why does AI-referred traffic convert better?
Mostly because of who is missing from it. Answer engines resolve informational needs in place and pass through what they cannot, which is disproportionately transactional, so the surviving visits are enriched for purchase intent before they reach you. A real causal component exists, but work separating the two puts it near 1.8x against a raw association of 5.7x. Treat any headline multiple as an upper bound, and never carry one into a forecast as a site-level assumption.
Key takeaway
A rising conversion rate is ambiguous evidence, equally consistent with a better website and a shrinking pre-qualified audience — and in most 2026 datasets the second does more of the work. Report it beside absolute conversions or not at all.
5. The four signatures
If rates are contaminated and counts are clean, diagnosis has to come from reading them together. Sessions, conversion rate and absolute conversions move in four characteristic patterns, each implying a different problem and a different budget response. Run it per page group, never site-wide, because the four routinely coexist inside one property.
THE FOUR SIGNATURES — reading sessions, rate and count together
Filtration. Sessions down, rate up, conversions flat. The engine absorbed demand that was never going to convert. You have not lost business; you have lost your measurement surface and your top-of-funnel influence window. The benign case, and the one that produces the most self-congratulation.
Displacement. Sessions down, rate up, conversions down. The engine is resolving intent that used to convert, or resolving it for somebody else. The expensive case, and the rising rate is camouflage — it keeps climbing while the business shrinks.
Absence. Sessions down, rate flat or falling, conversions down. Not a zero-click problem at all: you are not being retrieved, ranked or named, which has a different remedy and timeline. Diagnosing displacement when the truth is absence sends a team to rebuild measurement when it should be earning coverage.
Capture. Sessions flat or rising, rate up, conversions up. You are among the sources the answer draws on and receiving the residue as well. Rare, and worth dissecting when it happens, because it usually traces to a specific set of third-party sources rather than anything on your own site.
The transition is the point. Filtration and displacement look identical on a rate chart and opposite on a count chart. The first phase of being displaced looks exactly like the first phase of getting better, and it can run two or three quarters before the revenue line objects.
Why this cannot be done with rates alone
Conversions equal sessions times conversion rate. When both factors move for the same underlying reason, the product can sit still while the factors swing violently in opposite directions — and such a dataset supports two contradictory narratives, from which the analyst picks whichever suits the meeting.
The rule is blunt: in a zero-click regime every rate is contaminated and only counts are clean. Rates keep their diagnostic value and lose their reporting value, so any percentage in a board pack needs its count printed beside it.
6. What is left that you can actually forecast
Strip out everything the teardown marked amber or red and the position is uncomfortable. Query volume is no longer observable at the coverage a model needs. Citation rate is non-deterministic. Click-through is set by the completeness of an answer you never see. Three of the four inputs to the classic model are now properties of somebody else’s product roadmap. But a forecast need not be a prediction about the world; the useful ones are commitments about your own behaviour with a banded consequence attached, and that shape is still available.
Line one: committed supply
You can forecast, with ordinary project-management confidence, how many assets you will publish and how many independent third parties will carry an accurate, dated, retrievable statement about you by a given date. Both are production variables, with known lead times and a countable output that either exists on a given day or does not — the same logic behind treating your estate as anAI retrieval supply. This is the only term you can genuinely commit to, and it belongs on the forecast line.
Specify it as an operations plan specifies anything: quantity, type, target tier, lead time, failure allowance. Twenty-four placements across an agreed tier of trade titles, accreditation registers, independent editorial,sponsorship placements and sector roundups, at an assumed 60% conversion from outreach to publication, median twelve-week lead time. That is a forecastable object.
Line two: a banded conversion from supply to outcome
The conversion from supply to outcome is where the uncertainty goes, and it should come from your own history rather than avendor benchmark. Take the last four to eight quarters, count placements and assets shipped, count qualified enquiries in the quarter following, and produce a range. The band will be embarrassingly wide the first time. Publish it anyway: a wide honest band beats a narrow invented one, and it narrows as history lengthens.
Two disciplines make the band trustworthy. Lag it, because corroboration takes months to move retrieval and branded demand. And hold it to counts, because the moment a rate enters this line the contamination comes with it.
Line three: branded and direct demand, promoted
Resolved-but-not-arrived demand lands somewhere, and mostly it lands as a later branded search or a direct visit. One February 2026 dataset put 70.6% of confirmed AI traffic in the Direct bucket, because assistants strip or never set a referrer; Similarweb’s panel work found roughly 56% of AI-influenced visits misattributed as search. The classic model holds branded demand as a fixed baseline: somebody else’s win, constant while organic does the work.
In a rebuilt model it becomes a dependent variable: the destination of demand your visibility work created and never got credit for. Track branded search velocity against your supply plan, with paid brand spend held constant, and watch theassistant browsers that strip referrers entirely.
Be honest about this line, because it is the weakest. Attributing branded search growth to a supply plan is a correlation with a plausible mechanism, not a causal estimate. Log it as a hypothesis with a check date and a pre-committed falsifier. A model that quietly upgrades its shakiest term to load-bearing is how the last one got into trouble.
Key takeaway
The demand side became unforecastable. The supply side did not. Build the forecast on the term you can commit to — assets published, sources earned, by named dates — and express everything downstream as a band from your own history.
7. Why earned corroboration becomes the load-bearing line
An earned placement is a document. It sits at an address a stranger can load, published by a party that is not you, dated and stable in a way no output of a generative system is. Its existence does not depend on an engine’s cooperation this week, which is exactly why it can be committed to in a plan. A citation is an event inside a private, non-deterministic system, observable only by sampling. You can influence the second. You can schedule the first.
The supply matters as well as the schedulability. Muck Rack’s May 2026 analysis of more than 25 million cited links found earned media accounting for 84% of all AI citations; other 2026 breakdowns put brand-owned content at around 8% of ChatGPT citations against roughly 48% from editorial, forums, review sites and directories. When a retrieval step needs to answer a question your own site is structurally disqualified from answering — which suppliers exist in this category, how do these two compare, is this firm credible — it draws on third-party material or it draws on nothing. Your content estate is a stock you write once. The third-party layer is a flow, re-consulted from scratch every time somebody asks.
The placements a traffic model defunds
The placements that do most of the work in a retrieval-heavy environment frequently send no measurable referral traffic at all: an accreditation register, a trade-body directory, alocal citation or a sector listing that has sat at the same URL for six years. Nobody clicks them. They are consulted. A model whose dependent variable is sessions defunds every one of them with a clean conscience, because on its own terms they produce nothing. Judging a reference on its footfall is a category error your spreadsheet commits automatically unless the model has somewhere to put influence without arrival.
The routing maps onto the tactics you already run. Enumeration and shortlist questions are answered by directories, roundups and sector listings; comparison questions by independent editorial and earned commentary; credibility questions by dated mentions, accreditation registers and named expert quotes; category and market-context questions by the kind of entity-level corroboration that establishes what you are before anyone asks who is best at it. None of those live on your domain, and no amount of publishing puts them there.
Writing the line item
Put it in the model as production. Target count by tier and date. Assumed conversion from qualified prospect to published placement, from your own last four quarters rather than a guest posting or niche edit benchmark. Lead time by tier: registers and directories are fast and permanent, earned editorial is slow and prestigious, andlaunch surfaces sit between. A re-check cadence, because placements decay and one counted in January may not exist in June. And an exclusion list, because a model that pays by referring domain is a purchase order rather than a specification, and this publication has spent enough time on manual action recovery to know how that ends.
8. Worked example: Tarnbrook Cottages
Tarnbrook Cottages is an invented but specific case: a Lake District self-catering letting agency, 240 properties, £8.9M annual gross booking value, 22% commission, around 60% of bookings originating in organic search — a profile familiar to anyone runningmulti-region programmes. Its estate ran to 654 URLs, of which 412 were destination guides. The FY27 plan was built conventionally in September 2025: 41 keyword clusters, third-party volumes, a position-based click-through curve, a 2.1% session-to-enquiry rate, £1,180 average booking value. Baseline 61,000 organic sessions a month, forecast to 65,900 — 1,384 enquiries and 429 bookings a month, roughly £506,000 of monthly gross booking value. The board approved £340,000 against it.
January to March 2026: the good quarter that was not
In Q1 2026 sessions came in at 39,600 a month, 35% below baseline. The enquiry rate rose from 2.1% to 3.3%. Absolute enquiries landed at 1,307 — 6% under plan, 2% above the same quarter a year earlier.
The agency presented a quality story: traffic down, conversion up 57%, enquiries holding. The board accepted it, and was right to. Against the four signatures this is filtration. The guide library absorbed the hit — sessions to the 412 destination pages fell 71% — and those pages had always converted at around 0.4%. Booking and availability pages were down 9%.
April to June 2026: the signature flips
AI Mode became the default at Google I/O on 19 May, and Google’s own behavioural report that day noted planning-intent queries growing 80% faster than AI Mode overall — which, for a company selling three-night stays in Borrowdale, is the entire product.
Sessions fell again, to 31,200. The enquiry rate rose again, to 3.9%, the best number on the dashboard. Absolute enquiries fell to 1,217, 12% under plan and 5% below the prior year. Bookings fell to 371 a month against 429 planned.
The conversion chart hit an all-time high in the quarter the business started losing money. Filtration had become displacement, and the only reason anyone caught it inside eight weeks was that somebody had started printing counts beside rates in March. The revenue line would have taken two more quarters to make the same argument.
The teardown, run in June, produced three findings. Search Console showed 61% of clicks anonymised across the property and 79% on the guide library, so the forecast base was unobservable exactly where the estate was largest. Direct and branded search had risen 19% over nine months, and the old model booked all of it to brand, meaning to nobody. And of 40 tracked planning prompts across three engines, Tarnbrook was cited in 7, with 24 of the 33 misses naming an online travel agency — while the questions the engine did need answering, on availability windows, minimum stays and dog policies, were answered nowhere on the open web with authority.
July onwards: the rebuild
The model was respecified rather than recalibrated. Enquiries and bookings became the forecast line, in counts. Sessions moved to a diagnostics page with the four signatures printed against each of six page groups. The supply plan committed to 24 third-party placements across an agreed tier — two regional tourism boards, an accessibility accreditation register, four walking and outdoor titles, three dog-friendly travel publications, a national newspaper travel desk, sector roundups and eleven parish and valley-level local listings — plus 11 first-party assets, including a published dataset of minimum-stay and changeover-day rules across all 240 properties that did not exist anywhere else. The supply-to-outcome band came from their own eight-quarter history and was lagged one quarter. Branded search was promoted to a tracked dependent variable with a written twelve-month check.
At ten weeks: 9 of the 24 placements live against 14 planned; citation presence up from 7 to 13 of 40 prompts; branded search up 14%; enquiries recovered from 1,217 to 1,308 a month. Sessions fell again, to 30,100, and nobody escalated it.
The result that mattered was none of those. The rebuilt model’s Q3 band of 1,150 to 1,390 enquiries a month contained the actual. The old model’s point forecast of 1,384 had missed by more than 10% for two consecutive quarters. The rebuild’s success criterion was calibration, not growth — and calibration is what buys a marketing budget its next twelve months.
What went wrong
- The finance director rejected the band outright, on the grounds that a range is not a forecast. It took two quarters of the point forecast missing before it was accepted, grudgingly.
- Removing the conversion-rate chart from the board pack was the hardest internal argument of the year. It was the only slide that had been going up.
- Two committed placements — a tourism board listing and an accessibility register — carry nofollow links and produced 11 referral clicks between them in three months. They appeared in retrieval more consistently than anything else on the list, and both were proposed for cutting in the first budget review, under exactly the logic the rebuild existed to correct.
- Consolidating the guide library from 412 URLs to 180 cost more than 60 long-tail rankings and the largest referring-domain magnet the site had published. Rebuilding it as a dataset page recovered half those links in four months, and the decision is still arguable.
- The 40-prompt sample is small and non-deterministic: two runs a month apart differed by four citations with no work in between. And the branded-demand line remains unproven, because a brand campaign was running concurrently. The team recorded both rather than claiming the lift.
9. Where this argument could be wrong
The strongest objection is not that the data is wrong, but that the whole thing is over-thought.
Businesses have always modelled uncertain channels with rough bands and watched revenue. Selection effects in conversion rate are not new — branded search always converted better than non-branded, and nobody called that contamination. On this reading zero-click is an ordinary level shift dressed in the language of specification error, and firms that re-based their curves quarterly have done perfectly well.
That objection lands, and three parts are correct. Selection in conversion rate is old. Revenue is the ultimate settlement metric and always was. And plenty of firms rode this out without respecifying anything. Four things bound it.
The old selection effects were segmentable. Branded versus non-branded was a split you could make in Search Console and model separately, which is why nobody called it contamination. Engine filtration happens upstream of your measurement surface, so it cannot be split out and lands in the same bucket as your own performance.
A level shift converges; this does not. If the change were a one-off, quarterly re-estimation would close the gap and keep it closed. Instead the filtration boundary moves with product releases you never see: AI Mode by default in May, planning-intent growth, deep research modes, agentic checkout. Re-estimation against a moving boundary never converges; it lags.
Watch revenue is correct, and is not a plan. A lagging aggregate cannot allocate next year’s budget across content, earned coverage and paid. The intermediate model exists to decide where money goes before outcomes are known; deleting it defers that to whoever has the most confident spreadsheet in the room, which is almost always paid.
The comfortable period is the dangerous one. Filtration transfers value to you first — rates improve while counts hold — and takes it later, so the early phase of being displaced flatters precisely the metric you would use to conclude you are fine. Firms that recalibrated and watched revenue did well through filtration; that is not evidence the approach survives displacement.
What would falsify this
- If conversion rates for AI-referred and conventional organic traffic converge over the next several quarters while sessions keep falling, filtration is not selecting on intent and the contamination claim is wrong.
- If an engine publishes per-query citation and answer-completeness data to verified site owners, demand becomes observable again and the supply-side pivot loses most of its force.
- If organisations that only re-based their click-through curves show forecast error no worse than those that respecified, over two or more years, misspecification is a decorative distinction. This is the one that would hurt most, and it is measurable.
10. What to do on Monday
A literal sequence, and most of it is a week’s work.
- Print the counts next to every rate. Add absolute conversions beside every conversion-rate figure in your reporting. It costs an afternoon and it is the change that catches a flip two quarters early.
- Run the teardown on your own model. Sort every term into level-shifted, meaning-changed or discontinued, then count how many amber rows are forecast inputs.
- Read the four signatures per page group. Pull sessions, rate and absolute conversions for six page groups over six quarters and label each. Expect at least two different signatures in one property.
- Calculate your anonymised-click share. Property total clicks minus summed query-table clicks, divided by property total, per page group. Wherever it exceeds half, your keyword-based forecast base is a rumour.
- Pull your own click-through rates from Search Console, by question family. Not by position — position is not the unit any more, and a curve borrowed from another brand is worse than none.
- Build the supply line. Count placements and assets shipped per quarter for eight quarters, count qualified outcomes in the quarter following, and derive a lagged band. Commit next quarter’s supply by tier and date.
- Audit what your model would defund. List every earned placement producing under 20 referral clicks a quarter, then check which appear in retrieval. Anything on both lists is the asset a session-based model deletes.
- Set the re-check cadence. Archive every placement on acquisition and re-load the URLs quarterly. Budget the hours now; nobody budgets them later, and placements decay quietly.
The old model promised a number and delivered a story. The rebuilt one promises less: a commitment you can be held to, a band drawn from your own history, and counts that mean what they say. A smaller claim — and the first in three years that will survive being checked.
