Ask your agency a simple question: for the impressions you bought last month, who actually saw them? Not how many. Who. What kind of entity, on what kind of surface, under what kind of circumstances. In my experience, nobody in the chain can answer past the second follow up. The industry runs on a metric, the impression, that is structurally a claim made by the seller, verified mostly by companies the seller pays or tolerates.
That was survivable when the main pollutant was crude bot traffic. It is not survivable now. In 2026 the supply chain carries human views, headless browser views, AI agent views that represent real purchasing intent, AI crawler views that represent none, made for advertising pages whose only reader is an auditor, and synthetic surfaces generated faster than blocklists can name them. These are radically different things, worth radically different amounts, and they all clear the auction wearing the same costume: one impression.
I have come to believe the fix is not better fraud filters. Filters are a subtraction game on an adversarial field. The fix is provenance: treating every ad view like a shipment that must carry a chain of custody, and refusing to pay full price for cargo with no papers. Food went through this. Diamonds went through this. Attention is next.
Key Findings
- The impression has decomposed into at least five distinct species: verified human, probable human, commercial agent, non commercial machine, and fabricated. Only one deserves the historical price, and one deserves a price above it.
- Fraud detection asks "is this fake?" and loses the arms race; provenance asks "prove where this came from" and shifts the burden onto the seller, which is where economics says it belongs.
- What I call the Attention Chain of Custody can be assembled today from existing signals: supply path records, device attestation, interaction telemetry, and declared agent identity.
- Commercial AI agent views are not fraud; they are a new audience with real budgets behind them, and pricing them at zero is as wrong as pricing them as human.
- By 2029 I expect provenance graded inventory to trade in tiers, and unattested inventory to drift toward junk pricing, mirroring what unverifiable goods do in every mature commodity market.
The Impression Was Always Hearsay
Programmatic advertising is a marvel of engineering built on an evidentiary void. A bid request asserts a user, a page, a viewport. The advertiser's systems believe it, bid on it, and log the result as delivery. Nearly every party between budget and eyeball is paid on volume, which means nearly every party is paid to not look too closely. I walked through the plumbing in programmatic advertising explained, and the conclusion holds: the system audits itself the way a casino audits its own winnings.
Verification vendors improved things at the margins, but they inherited the structural flaw: they sample, they estimate, and they answer the subtraction question. Was this impression invalid? The adversary only needs to be slightly more creative than the taxonomy. Meanwhile a new complication arrived that taxonomies were never built for: machines that view ads while doing legitimate work.
The Five Species of a View
Here is the classification I now audit against.
Verified human: a view with strong evidence of a person, on real hardware, with the ad rendered in a viewable state and some sign of engagement or at least presence. This is the species every historical benchmark assumed.
Probable human: consistent signals, no strong attestation. Most decent inventory lives here. It deserves a modest discount to verified, not parity.
Commercial agent: an AI system browsing, comparing, or purchasing on behalf of a real customer. It ignores your banner but reads your specifications, your pricing, your reviews. As I argued in marketing to AI agents, this species carries genuine economic intent. It is not an impression in the old sense, but it is absolutely a marketing contact, and in some categories it will become the dominant one.
Non commercial machine: crawlers, scrapers, model training fetches, monitoring tools. Zero persuasion value. Preparing your infrastructure to identify and manage this species is work I described in preparing your website for AI crawlers.
Fabricated: traffic that exists to be sold. Botnets, device farms, spoofed apps, made for advertising chains. Negative value, because it also poisons your optimization data and teaches your algorithms to hunt more of it.
The scandal of current measurement is not that the last species exists. It is that the ledger records all five as the same unit.
The Attention Chain of Custody
Provenance means that by the time a view is billed, it carries an unbroken, inspectable record: which seller originated the request, through which resold hops it traveled, what surface rendered it, what attestation the device or environment offered, what the interaction telemetry looked like, and whether the viewer declared itself an agent. Call it the Attention Chain of Custody. None of the components are science fiction. Supply path transparency standards exist. Device attestation exists. Agent identity declaration is emerging as agents become commercially important and want to be served well. What is missing is the buyer's demand that the pieces be assembled and the willingness to pay differently based on what they show.
The economic logic matters more than the technology. Under a filtering regime, the buyer bears the cost of detecting lies. Under a provenance regime, the seller bears the cost of proving truth, and inventory that cannot prove anything simply trades at the price uncertainty deserves. You do not have to catch anyone. You reprice the silence.
Run the illustrative math. Suppose, and I am inventing these numbers to show the mechanism, a campaign's million impressions decompose into forty percent verified human, thirty five percent probable, ten percent commercial agent, ten percent non commercial machine, and five percent fabricated. A blended CPM prices all million identically. A provenance graded buy might pay a premium on the forty, standard on the thirty five, a small strategic rate on the agent slice, and zero on the rest. Same budget, radically different distribution of dollars toward things that can actually move revenue. That reallocation, repeated across the market, is how junk supply finally starves.
Agents Are an Audience, Not an Anomaly
I want to press on the species the industry is mishandling worst. The reflex is to classify all machine views as waste. But an agent dispatched by a household to research a purchase is carrying delegated intent and delegated money. Blocking it, or ignoring it, is refusing a customer because they sent a representative.
The right posture is separation, then strategy. Identify agent traffic honestly, exclude it from human reach metrics so your brand numbers stay clean, and then serve it deliberately: structured data, machine readable claims, transparent pricing, verifiable reputation. In provenance terms, the agent species should get its own line in every report and, eventually, its own budget. My prediction: by 2030, sophisticated advertisers will report human reach and agent reach as distinct KPIs with distinct targets, and the ones who merged them will spend two years untangling their attribution.
The Market That Grades Itself
Falsifiable forecasts, on the record. By 2028, at least one major exchange or buying platform offers explicitly tiered inventory keyed to provenance strength, priced visibly apart. By 2029, unattested long tail inventory trades at a steep structural discount, and made for advertising supply collapses because its only product was ambiguity. By 2031, provenance reporting appears in the standard quarterly deck of any serious CMO, the way viewability crept in a decade earlier. If instead the blended impression is still the universal clearing unit in 2031, I got this wrong.
I do not expect regulators to lead this. I expect finance departments to. The first CFO who demands a chain of custody for a hundred million dollar media line will do more for clean supply than a decade of industry working groups.
What I Would Do About It
Start with an audit you control. Take your own server logs and analytics for one quarter and classify inbound traffic into the five species as best current tooling allows. Your own property is the one place you can see ground truth, and the decomposition will calibrate your skepticism about everything you buy elsewhere.
Rewrite your buying terms. Ask every major supply partner one contractual question: what provenance evidence accompanies each billed view, and what is the discount schedule when evidence is absent? Their discomfort is information. Concentrate spend with the partners who answer in writing.
Separate your species in reporting immediately, even crudely. A dashboard with verified human, probable, agent, machine, and unknown columns changes internal conversations within a quarter, because it makes the ambiguity everyone profits from visible to the people paying for it.
Build your agent lane. Treat commercial agent contact as a channel with its own content requirements and success metrics rather than noise to filter.
And push your optimization systems to train only on provenance clean outcomes. Algorithms fed fabricated signals develop fabricated appetites. Clean the diet before you scale the spend. The advertisers who demand papers for their attention will, within five years, be buying a different and better product than the ones who kept paying for costumes.