What the gate admits, and who loses

1 · The diagnosis: what the current auction actually sells

A search auction of the conventional kind does not sell relevance. It sells proximity to intent, and it prices relevance as a discount rather than a condition. Because the quality estimate multiplies the bid instead of gating it, a sufficiently large bid always clears a mediocre relevance score. Every pathology below follows from that one arithmetic choice.

The second choice compounds it. Per-click pricing aligns the platform's revenue with the click, not with the resolution of the question. A user who is confident does not click. A user who is uncertain clicks several times. Under CPC, uncertainty is inventory — and the platform has no mechanism, and therefore no incentive, to end the session early with a correct answer.

The three pathologies are not abuses of the model. They are the model working correctly: brand conquesting monetises someone else's demand, lead arbitrage monetises a customer's inability to find the supplier directly, and harm-adjacent placement monetises a bad moment. All three are rational purchases of proximity.

2 · Inventory we expect to admit

The gate admits a bidder only where their own published claim answers the classified question. That makes the viable inventory narrow and predictable: questions with a factual, checkable answer where a named product is a legitimate part of it.

Table 1. Expected admissible inventory. Note that every high-value class is a question the reader could not answer from the answer text alone — a specific vendor is part of the resolution, not a substitute for it.

3 · Inventory we expect to be offered, and refuse

We should expect every category below to arrive within the first quarter, several of them with large budgets and competent representation. The mechanism refuses most of them without a decision being taken; two require an explicit rule.

Table 2. Mechanism = refused by arithmetic, no judgement required. Review = refused by a human checking claims against live pages. Rule = requires an explicit prohibition; only two categories do.

4 · Brand conquesting: the case that matters most

Bidding on a competitor's brand name is the single largest behavioural difference between Google and Claim #1, and it deserves a precise answer rather than a slogan. The gate does not ask "is this a competitor?" — it asks whether the bidder's own claim answers the question that was asked. That distinction splits conquesting cleanly in two.

Table 3. Conquesting dies on navigational intent and survives on comparison intent — which is, we would argue, the correct division. A reader asking for alternatives is asking to be shown competitors.

The consequence for incumbents is larger than the consequence for challengers. A mature search programme spends a substantial share of its budget bidding defensively on its own brand, for no reason other than that a competitor would otherwise occupy that slot. Under Claim #1 there is no slot on brand-navigational intent for anyone, including the brand owner — so the defensive line item does not shrink, it ceases to exist. That is the clearest and most quantifiable saving we can offer an enterprise buyer, and it is worth more to them than anything they can buy from us.

5 · Gambling, and the one place our list is wrong

The original paper's prohibited list covered health, legal and financial advice, elections and crisis. The live pilot now also blocks gambling, betting, loot boxes and trading signals. The argument is not moral; it is a limitation of safe intent separation.

The current pilot applies deterministic hard-stop phrases and then matches question terms against the advertiser's published summary. It does not claim that a trained classifier separates every harmful and commercial intent safely. Gambling illustrates why the hard stop remains necessary. "Best odds on tonight's match", "how do free bet offers work" and "how do I stop gambling" occupy adjacent regions of the same intent space, share vocabulary, and are asked by overlapping populations, sometimes by the same person within one session. A classifier that admits the first will, at some rate we cannot drive to zero, admit the third.

Two secondary observations reinforce it. First, the economics of gambling acquisition depend on per-click volume, cross-session retargeting and urgency language — all three of which our mechanism already removes, so the category has little to gain here even if permitted. Second, publishing the amount is uniquely hostile to affiliate arbitrage: an affiliate whose margin is the difference between a public claim price and a private revenue share has that margin rendered legible to the operator they resell. The category will not want us. We should nonetheless say no in writing first, because the alternative is saying no later, under revenue pressure, having already accepted a payment.

6 · Who wins, who loses

Table 4. The uncomfortable row is the agency one. We remove bid management, negative keywords, quality-score work and creative testing in a single stroke, and we should expect the channel that sells those services to advise against us.

7 · The honest cost to the advertiser

A Google buyer moving to Claim #1 experiences a collapse in volume and a very large rise in unit price. On the 30-day worked example, a bidder appears on 26 of 41,200 board answers — about 0.06% — and the derived effective cost per click is $307. The annualised equivalent is 312 of 494,400 answers at $9,000. Against enterprise-software keywords that already clear $50–$100 per click this is a three-to-six-fold increase, not the hundred-fold it appears against consumer averages, but it is an increase and we should quote it that way.

Three offsets are real and one is not. The real ones: the budget becomes a fixed, prepaid, forecastable cost rather than a variable one, which is materially easier to defend to a finance function; the eliminated defensive brand spend is pure recovered margin; and click quality rises because the reader arrives having already read a correct answer rather than in search of one. The offset that is not real is reach. We should never imply that a gated slot reaches more people. It reaches far fewer, and that is the design.

8 · Conclusions

The problem was never disclosure. It was that relevance was priced as a discount instead of enforced as a condition, and that the platform was paid for the click rather than for the resolution. Every objectionable format in search advertising is a rational response to those two facts.

The change is arithmetic, not editorial. A gate that does not take the amount as an input cannot be bought past. A flat price paid before any click removes our interest in the click, and therefore our interest in the reader's uncertainty. Placement after the seal removes the answer from the market altogether. We did not prohibit conquesting or arbitrage; we built a mechanism in which they have nothing to purchase.

The inventory is small and should stay small. Sixty-nine per cent of answers carry no slot at all. The temptation, once revenue is real, will be to widen the gate by a few points of confidence and recover that inventory. That single change would undo everything above, and it is the one number an external auditor should be watching.

Our most valuable product is a thing we refuse to sell. Brand-navigational intent carries no slot for anybody. That is simultaneously the largest revenue line we are declining and the strongest argument an enterprise buyer will ever hear from us, and the two facts are the same fact.

Companion to Working Paper 2026-04, Advertising without entering the answer. Figures are illustrative; budget shares vary by sector. Correspondence: research@agicy.ai · agicy.ai/research · September 2026