AI IN MARKETING WEEKLY

Issue · September 14, 2026 · AI Search + Measurement

Ad buyers just made getting found inside AI answers their number one priority. A paper posted six days earlier says one check shows you two-thirds of the picture.

The IAB asked more than 200 brand and agency ad investment decision-makers where their focus is growing fastest. The top answer, at 76%, was optimizing content for AI-generated answers. It beat every other option on the list, and focus on using generative AI inside campaigns actually fell, from 78% in January to 69% now. Eighty-six percent say conversational AI is changing how they measure media. Forty-eight percent have landed on how.

By Ged King · Sources: IAB, “IAB Raises 2026 U.S. Ad Spend Forecast,” September 10, 2026 (primary) · Żatuchin, arXiv:2609.05059, September 4, 2026 (primary, preprint)

Start with the forecast, because it's the good news. The IAB raised its 2026 U.S. ad growth number to 12.3%, up from the 9.5% it called in January. Social got revised up to 16.5%, connected TV to 15.6%, commerce media to 13.6%. Paid search got revised down, to 8.1%, which makes it the second-slowest growing digital line in the IAB's own table. Linear TV is still negative.

Now the part worth your morning. In January this industry's biggest question was how to use generative AI in campaigns. Eight months later that focus has dropped nine points while “how do I get found by it” jumped to the top of the list. Forty-four percent of buyers say adapting to AI-driven search is now their single biggest media investment challenge. Thirty-eight percent say they're worried about low-quality AI content polluting the places they buy.

So a lot of budget is about to go looking for an AI visibility tool. Which brings me to a paper that went up on arXiv on September 4, and I'd read it before you sign anything.

The setup is simple. Fifty questions. Six AI engines. Fifteen runs of each question on each engine, 300 question-engine cells in total, with 1,470 organizations pulled out of the answers. The question being tested: ask a model the same thing over and over, does it eventually run out of brands to name?

For five of the six engines, no. The five answering without web search were still producing brands they had never mentioned before on the fifteenth run, in 86 to 92 percent of cells. Only the one engine with retrieval turned on settled down, and even that one was still adding names in 64 percent of cells. The number to write down: a single run shows 62 to 77 percent of the brand set you'd see across five runs.

Then the author ran a control, and the control is the part I keep coming back to. He scored the exact same model responses a second way, against a fixed roster of brands instead of extracting whatever names appeared. The fixed-roster version produced flat, stable curves. Same responses. The stability came from the method, not the model.

Most AI visibility tools score you against a preset roster of competitors. That isn't a scandal, it's how they're built. But it does mean the reassuring flat line on the dashboard may be describing the list somebody typed in during onboarding. And I'll be straight with you about the source: this is a single-author preprint, it hasn't been peer reviewed, fifty questions is a small corpus, and the engines aren't named. Weigh it accordingly. The control experiment still holds up better than anything else I've read on this all year.

What it means Send your AI visibility vendor two questions this morning. How many runs per prompt? And open extraction or fixed roster? If they run each prompt once against a preset competitor list, you're buying a chart, not a measurement. A companion paper from the same author puts the bar for anything you'd set budget on at fifteen runs, while noting honestly that those tiers didn't transfer cleanly to independent data. And if you're the one reporting AI visibility upward, this is your cover for saying “directionally” instead of putting a decimal point on it.