Sponsored answers in AI assistants, bought against holdouts

An impression inside a generated answer is not a page impression, and the platform's own report will not tell you the difference. On the assistant placements I have run, sponsored follow-up questions on Perplexity, Microsoft Advertising inventory surfacing inside Copilot, sponsored results in Amazon's shopping assistant, measured incremental conversion has landed between a half and a quarter of the platform-reported figure. The buy is only defensible if the holdout exists before the first euro is spent. What is actually being bought: Not keywords. The unit is a question a user asks and the moment the assistant offers a next step. The brief opens with a question taxonomy for your category, first-time intent, comparison, objection, troubleshooting, repurchase, written from your support tickets and site-search logs rather than a keyword tool. Placements are bought against that taxonomy, and inventory where the platform will not say which questions triggered delivery is left unbought. Creative that survives paraphrase: The assistant rewrites you. Copy carried by tone, rhythm or a slogan disappears in that rewrite. A single checkable claim with a number and a named product survives it and gets repeated. The source pages behind the placement are written the same way, one claim per paragraph, because the same model reads them. This is what most media teams get wrong on the first flight, and it is testable inside three weeks. Measurement: A geographic or audience holdout sized to your conversion volume, a test window of four to six weeks, and an incrementality read rather than a platform-attributed one. Post-click behaviour is tracked to the same conversion event you already use elsewhere, not to a channel-specific proxy that flatters the channel. Brand lift is reported only when the sample supports it, which at this spend level it usually does not. Not included: Organic visibility work. Being cited in answers you have not paid for is a different discipline and I do not sell the two together, because the same person owning both has an obvious incentive to confuse the reporting. No landing page development, no PR, and no work on standard search campaigns, which your existing agency runs and should keep running. Who this is not for: Buyers who need a last-click ROAS number in a weekly deck. This channel loses that argument every time and I will not adjust the attribution to help it win. Budgets under roughly 25,000 euro a quarter, because a holdout at lower volume returns a confidence interval wide enough to support any conclusion you like. Categories the assistants decline commercially anyway, regulated finance and prescription health among them. How the engagement runs: Two weeks to build the question taxonomy and the measurement design, then a first flight, then a decision that is genuinely allowed to be no. I have recommended stopping after one flight twice. Both times the incremental read sat inside the noise band, and the honest answer was that the inventory did not yet justify the management overhead. Reporting arrives as a written read with the raw export attached, addressed to whoever has to defend the line item internally rather than to whoever signed the insertion order. Working language is English. Markets run to date: Ireland, the UK, the US and the wider EEA.