Original insights by Limor Barenholtz
30-second rundown
Key learnings:
- AI influence hides behind other channels: only 8.8% of visits following an AI recommendation arrived through a channel labeled AI, while most appeared as search, direct or another referral.
- Evidence pages and arrival pages need different jobs: protect deep pages that earn citations, even when buyers later enter through the homepage.
- Five honest measures beat one false attribution story: track Mentions, cited Evidence, Traffic, Elapsed time and buyer Recall (METER) separately.
This week: Compare AI mentions with branded search and landing-page traffic for the same recent period, and record the five METER measures in a simple monthly scorecard.
Coming 3 months: Assign one owner to an automated monthly workflow that captures AI mentions and citations, branded-search movement, landing pages, time-to-visit signals and self-reported discovery. Review the five trends side by side so budget decisions improve without pretending each touchpoint belongs to one traceable person.
Limor Barenholtz has found the corpse of tidy marketing attribution beside a sun-bleached highway, pockets turned out and fingerprints everywhere. AI did not destroy the buyer journey. It split the evidence across systems that cannot recognize the same buyer, then let every dashboard swear it saw the whole crime.
1. The click trail no longer tells the story
Barenholtz draws a useful line between a dark funnel and a fragmented one. A dark funnel contains activity nobody recorded, such as a private message. A fragmented funnel records the pieces but cannot prove they belong to the same person, like five witnesses describing one getaway car in five different colors.
Similarweb tracked thousands of US desktop journeys in finance, travel and beauty. People receiving a ChatGPT brand recommendation were 2.5 times more likely to visit the recommended brand than a competitor within seven days. Yet only 8.8% arrived through an AI referral. Search delivered 55.9%, direct 19.9% and other referrals 13.5%.
That matters because a dashboard judging AI by referral traffic sees roughly one-tenth of the measured effect. The rest gets credited to branded search, direct traffic or some other channel wearing a clean suit and a guilty smile.
2. The page that supplies proof may never receive the visitor
The fracture runs inside the website. Similarweb found that 41.7% of cited URLs sat two folders deep, while 58.8% of AI-referred visits landed on homepages. The page that convinces the model is often not the page that welcomes the buyer.
Picture a detailed comparison page with forty citations and almost no direct visits. A traffic-only content audit sees a dead page and reaches for the shovel. In reality, that page may be the evidence AI uses before the buyer searches the brand and enters through the homepage. Delete it and the damage may surface weeks later as weaker mentions, lower branded demand and a baffled finance meeting.
Barenholtz recommends separating proof pages from arrival pages. Measure proof pages by citations and influence. Measure arrival pages by their ability to explain the offer and convert a buyer who may already have spent twenty minutes interrogating an AI system.
This is not an argument for preserving every dusty article. Add citation data before pruning. Traffic, citations and conversion readiness are different gauges. Smashing them into one number is how the car catches fire.
3. METER replaces the fantasy with five useful numbers
The proposed METER framework tracks Mentions, Evidence, Traffic, Elapsed time and Recall. Mentions ask whether the brand appears in relevant AI answers. Evidence records which URLs are cited. Traffic shows where people arrive. Elapsed time measures the delay between AI influence and later behavior. Recall captures what buyers say brought them to the website.
The governing rule is wonderfully severe: never add the five measures together. They have different denominators and often describe different populations. A single combined score would look precise while inventing a relationship the data cannot establish.
Instead, read the pattern. Mentions rising while citations stay flat can mean the brand is discussed without first-party evidence. Citations rising while traffic stays flat can mean proof pages are working but the brand is not memorable enough to earn the later visit. A longer delay may reflect a more considered purchase rather than marketing failure.
For a small business, track a stable set of buyer questions, cited pages, branded-search direction, entry pages and a “How did you hear about us?” field. A spreadsheet and recurring review reveal more truth than a model claiming to reconstruct people it cannot identify.
The old funnel promised a neat arrow from first touch to sale. That arrow now lies rusting behind the motel, and good riddance. Keep the five receipts, measure each for what it actually proves, and stop making budget decisions with counterfeit certainty.
