Nikhil Lai covers performance marketing and advertising as a principal analyst at Forrester. On a recent Marketecture episode, the conversation reached Mobian's ads inside Time's machine-readable pages, the same practice Perplexity is now docking publishers' trust scores over. The hosts called agent-targeted ad formats inevitable.

Lai said Forrester is publishing predictions in the next couple of months on the rise of non-human traffic and how to persuade non-human traffic to recommend you.

Not tolerate it. Not filter it.

Persuade it.

He was explicit that this is not next-generation SEO but a whole new skill set marketers have to build.

Then came the part everyone will scroll past, and it is the only part that matters.

He asked what KPI applies. Viewability? Opportunity to see? Attention? He concluded none of them fit and that a new one is coming, then named it: recommendation share. A mid-funnel cousin of share of voice, which he expects to correlate with lower-funnel outcomes.

So an analyst at the firm that helped teach this industry to treat non-human traffic as fraud looked straight at the measurement problem, found that no existing metric survives contact with it, and invented a new unit of account instead of answering whether the old ones apply.

Understand exactly what got skipped there.

Viewability, opportunity to see and attention are counted-media metrics. They live inside a system with definitions, auditors, an accreditation body, and a filtration layer whose entire purpose is to strip this traffic out before anyone gets billed. Asking whether agent exposure is viewable is asking whether it is an impression. And if it is an impression, the existing rule already answers it: bots, spiders, crawlers, non-browser user agents and prefetch activity are general invalid traffic when they generate associated media activity.

Recommendation share walks around all of that. Earned, not bought. Off-site, not on-page. Mid-funnel, not billable. No auditor, because no definition. No definition, because it is a few weeks old and was born on a podcast.

Credit where it belongs. Lai is being more honest than most of the people who are going to monetize this. He is not claiming agent impressions should be billable. He is arguing that presence inside an answer is a different object from a served ad, and on that narrow point he is right.

But watch what happens the second you accept it. The measurement question is not resolved. It is relocated to a jurisdiction where nobody has authority.

While the metric was being named, the largest retail ad platform on earth was telling a judge what it thinks of this traffic.

Suing over Perplexity's Comet browser, Amazon argued that AI-generated ad traffic has to be detected and filtered before advertisers can be charged, and that it needed to build new detection to do it. That is not a position paper. It is the existing rule with a case number attached.

August 2026, then. Amazon is paying engineers to scrub agent traffic off the invoice. Forrester is drafting guidance on how to attract it. Both are internally coherent. Neither has ever been reconciled with the other in public.

Same machines. Two invoices. One reads fraud. One reads funnel.

This is not one analyst freelancing. Forrester's own position has shifted, and the moves have dates on them.

In 2014, a Forrester senior analyst was presenting non-human traffic to advertisers under the heading of digital advertising's dirty little secret, with botnets, ad stacking and impression laundering as the exhibits.

In 2024, the Forrester wave for this market was called Bot Management Software, and its findings led with bad bots getting more sophisticated and more available.

This year, the wave is called Bot And Agent Trust Management Software. The vendors in it read the rename as formal recognition that the market has moved past blocking bots toward managing trust across humans, bots and agents.

And in March, Forrester published a post arguing that marketers historically threw bot traffic away as noise or fraud, that the reflex made sense when crawlers and criminals were most of it, and that this world is over. Its position: user-initiated agents fired by a real buyer's prompt are upstream intent, and discarding them means discarding your earliest evidence of influence.

A research firm renaming a category is what research firms do instead of admitting they changed their minds. The old name assumed automation was a threat surface. The new one assumes automation is a constituency you negotiate with. Whether that constituency's ad exposures are billable is left, as ever, as an exercise for the reader.

Suppose you take the advice, execute well, and machines start recommending you constantly.

Analytics suppresses most of it as bot traffic. Your CDN logs show user agent strings any competent party can forge. Your verification vendor cannot see inside an answer engine and is contractually aimed the other way. Your attribution model has no path from an answer nobody clicked to a purchase three weeks later.

Lai is honest about this part too. Nobody has connected view-through traffic from answer engines to lower-funnel outcomes, and he called that the next frontier. His own proposed workaround is to baseline branded and direct search, watch for a spike, and attribute a slice of it to answer engine exposure. He called it speculative himself. It is econometrics performed on a rumor.

Which leaves this business somewhere it has never been. The only available proof of performance in the fastest-growing channel in marketing is asking the model whether it likes you. That is not measurement. It is a focus group with one respondent, who is also the media owner, who will not publish the methodology, and who spent this month assigning publishers secret trust scores.

There is a version of recommendation share that becomes a real metric. It needs a published definition, an independent auditor, a stated relationship to counted media, and a rule for what happens when the session that produces a recommendation also loads a page full of ads.

There is another version, and it is the one this industry builds every single time. Name the metric. Sell the metric. Accredit it in four years, maybe, for one platform, in one geography, if anybody remembers to ask.

I have watched the second version happen roughly a dozen times on this beat. Every gap between what a metric claimed and what anyone actually verified got filled by a vendor with a deck and a logo slide.

So the question is not a gotcha, and it is not aimed at Lai, who raised the hard part out loud when he could have simply sold the upside.

Is recommendation share going to have a definition and an auditor, or is it going to have a sales team?

The guidance publishes in a couple of months. The vendors are already building against it. The invoices are already going out twice.

How we reported this: sourced from a transcript of the Marketecture episode quoted here, Forrester's published post on zero-click buyer data in bot traffic, a 2014 Forrester analyst presentation on non-human traffic, MRC invalid traffic definitions, and Amazon's filings against Perplexity. The transcript is machine-generated and garbles names throughout, so quotations and spellings were verified against the audio. Forrester wave titles and dates were reviewed through vendor reprint pages rather than licensed copies. Forrester was contacted about whether it holds a position on the billability of agent-generated ad exposures and whether recommendation share will carry a published definition. ADOTAT has no financial relationship with any company named here, no subject reviewed this before publication, and there is no paid placement in editorial.