Here is how your Black Friday report is going to lie to you. Somebody asks an AI assistant for the best carry-on bag under two hundred dollars. The assistant names three, yours is one of them, and the shopper closes the chat without clicking anything. Four days later, on the Friday, they type your brand into Google, click your branded ad and buy. Your dashboard records a branded search conversion at a lovely cost per sale. In January someone looks at that number and asks for more branded search budget.
The ad did not create that sale. It caught it. And on November 12 the Interactive Advertising Bureau published a framework that says, in plain language, this is now one of the main ways marketing money gets sent to the wrong place.
The document is called the AI-Era Attribution Blueprint. It will not fix your measurement. It does not hand out percentages of credit, it names no auditor and nobody is obliged to follow it. What it does is give you a vocabulary for an argument you are going to have in about seven weeks, and a reason to set the rules of that argument before the sale, not after.
What the IAB actually published
The blueprint sorts every sale by three questions. Who or what created the interest? Who or what made the decision? Who or what completed the purchase? The answers put each conversion on one of three paths.
Human-Owned means a person did all three, even if an assistant answered a side question along the way. AI-Assisted means a person bought, but AI shaped the interest, the choice or both. The IAB splits that into AI-Influenced, where the assistant put the idea in someone’s head and they decided, and AI-Recommended, where the assistant effectively picked the product and the person accepted. AI-Executed means a person gave an agent permission ahead of time and the agent completed the purchase itself.
Then comes the line every media seller will try to skip past. The blueprint says a claim is not credit. Being on a path only makes a party eligible to be evaluated for credit, and every claim needs evidence. According to PPC Land’s walkthrough of the 25-page document, the IAB refuses to assign fixed shares of credit at all, because there is no defensible basis for splitting credit in journeys that leave partial trails or none.
Caroline Giegerich, the IAB’s vice president for AI, put the reason for all this bluntly in the announcement: AI-influenced sales “are sometimes reported as search or direct traffic, which can send the next marketing investment to the wrong channel.” Speaking on an IAB webinar, she told AdExchanger the categories are built on “what’s actually happening,” not on what anyone can measure today.
That second point matters more than it sounds. If the categories were defined by what your tools can see, every AI-shaped sale without a surviving referral would quietly default back to search or direct, and the problem would vanish from the spreadsheet while staying very real in the market.
Where the AI sales are hiding
The scale argument rests on outside data. In a study published in June, Similarweb found that people who got a brand recommendation from an AI assistant were 2.5 times more likely to visit that brand’s site, and that 55.9% of those AI-influenced visits arrived through search. The blueprint, as PPC Land reports it, adds the rest of the split: 19.9% arrived direct, and only 8.8% came from the AI platform itself. AdExchanger summed up the same Similarweb finding as more than 90% of AI-driven sales being miscredited as branded search, direct or retail media.
The IAB also says its own buyer survey shows the industry already knows this and has no better tool. In its 2026 Outlook Study September update, according to the announcement and MarTech’s write-up, 45% of ad buyers named comparing AI-driven and traditional journeys as their top measurement challenge, and 44% said they lean on branded search and direct traffic as stand-ins for AI influence they cannot track.
Read that again from a budget seat. Almost half of buyers are using the very channels that soak up AI’s work as the proxy for AI’s work. That is fine as a rough signal. It is a disaster as a budget rule, because the proxy and the channel are the same line item. Branded search goes up, so branded search gets more money, so branded search goes up.
The blueprint has a second scenario that should worry anyone running retail media. A shopper researches in an assistant, arrives at a retailer with a short list, sees the retailer’s sponsored listings and buys. The retail media report shows a tidy closed loop. But the assistant may have decided which brands made the list in the first place, and current systems cannot tell whether the ad created demand, swayed the final pick or just stood next to a decision already made.

Why this holiday tilts even harder toward invisible
The easiest AI sales to measure are the ones an agent completes, because an agent checkout leaves a record: the agent identifies itself, the merchant confirms the order, the payment carries the shopper’s permission. That is the path the blueprint can see best. It is also the path big sellers spent this week closing.
On November 11 United Airlines revised its Contract of Carriage to bar AI agents from making, changing or servicing tickets or buying products, and to list agent bookings among reservations it may cancel, whether or not the passenger supervised or approved what the software did. United’s announcement cites account security and regulatory rules, and Search Engine Journal notes that United still lets search engines index its public pages while requiring written permission for agents. The same week The Verge reported that Meta’s Muse agent had stopped reliably reaching eBay listings, and PPC Land pointed out that eBay’s user agreement has barred buy-for-me agents without its permission since February. eBay told The Verge its restrictions are not a blanket ban.
I am not making a prediction about agents here. The point is narrower. If the measurable route, agent does the buying, is being fenced off by the largest sellers, then whatever AI does for your brand this season will mostly travel the route that leaves no trail. Somebody asks, somebody reads, somebody later searches your name. That sale will look exactly like a branded search win.
What the blueprint will not do for you
It is worth being honest about the limits, because the vendors pitching you in December will not be.
It is voluntary. It sets no threshold for when evidence is good enough. PPC Land notes that none of OpenAI, Google, Amazon, Meta, Perplexity or Anthropic appears among the named contributors, even though most of the disclosures it asks for would have to come from companies like them. Those requests are specific: tell advertisers whether a product appeared because of a paid placement or organically, say what sources informed a recommendation, and label AI-originated traffic consistently. None of that exists in a usable form today.
It also treats AI visibility scores, the presence, rank and share of prompt numbers that a fast-growing tool category sells, as directional indicators rather than proof. That is a useful sentence to have in your pocket when someone shows you a share of voice chart and asks for budget.
And it is clear that the only real evidence for the influence path is the same evidence TV has always needed: compare people who were exposed with people who were not. Paid AI placements can be withheld from a control group. Organic answers mostly cannot. Surveys and modeling help, but the IAB calls them directional.
So the framework does not give you a number. It gives you permission to stop pretending the number you have is the answer.
The decision this week
The useful move is not measurement heroics. It is writing down, before November 27, how you will read November 27. Once the sale is over, every team will look at the same spike and find a reason it was theirs. A rule written in advance is the only kind that survives that meeting.

What to do before Black Friday
Write the January rule now. One paragraph, signed by whoever owns the budget: a holiday lift in branded search or direct traffic does not move money into branded search on its own. It moves money only after a holdout, a brand lift read or at least a survey cut says the demand was created by paid media and not caught by it.
Add one question to your post-purchase survey. If you already ask “how did you hear about us,” add “did you use an AI assistant like ChatGPT, Gemini or Copilot while deciding what to buy,” with options for yes, it suggested you; yes, I asked about you; and no. It is directional, as the IAB says. It is also the cheapest evidence you can collect, and the only kind that covers sales with no trail at all. It has to be live before the sale or you get nothing for this season.
Take a baseline this weekend. Snapshot branded query volume, branded cost per click, direct traffic share and new versus returning buyers for the last four weeks. When January arrives, a branded spike with flat new customer share tells a different story than a branded spike full of first-time buyers.
Hold something out where you control the switch. If you are running paid placements inside an AI assistant, keep a geography or audience dark for the season. That is the one AI exposure you can actually withhold, and the blueprint is explicit that a comparison like that is what turns a claim into evidence.
Ask every claim three questions. When a platform, agency or AI visibility vendor says AI drove a share of your holiday sales, ask which path the sales were on, what kind of claim is being made, and whether the evidence is their own reporting or something you can check independently. Keep platform-reported numbers in a separate column from verified ones in the January deck.
Ask your retail media partner the awkward question. If a retail network reports strong closed-loop sales, ask whether it can show any of that demand originated off its site, and what share of buyers arrived with your product already on their list. You probably will not get an answer. Asking now still changes how you read their recap.
Pull AI-executed orders out if you can see them. If your platform or checkout records agent orders, tag them so they do not blend into human conversions. The blueprint says an agent following a specific instruction gets no claim, it is a service, so do not let a vendor count those reorders as AI-driven growth either.
None of this is glamorous. That is the point. The biggest risk this season is not missing AI’s contribution. It is rewarding the channel that happened to be standing at the till when the shopper showed up, then paying it more all year for a sale it caught.




