Twelve months ago, the AI referral line in a retail analytics stack was the worst performing thing in it. Shoppers who arrived from an assistant bought less often than shoppers from paid search, email, or affiliates. Today the same line is the best performing thing in it.
In March 2026, traffic from AI sources to US retail sites converted 42% better than non-AI traffic, a record high. In March 2025, that same traffic converted 38% worse.
Conversion of AI-referred traffic against non-AI traffic, US retail. The two figures are measured against different baselines; the direction is the finding.
Those two figures are measured against different baselines, so the tidy “eighty point swing” being quoted around this report is looser than it sounds. The direction is not loose at all. In twelve months a channel went from the worst performer in US retail to the best of them, and the volume came with it. AI referred traffic grew 393% year over year in the first quarter of 2026, and those visitors stayed 48% longer and generated 37% more revenue per visit than everybody else.
Retail is where this shows up first, because retail transacts fast enough to measure inside a quarter. Nobody publishes a comparable dataset for a nine-month enterprise sale. The behavior underneath it is the same behavior, and in a considered purchase it is harder to see and costlier to miss. McKinsey puts $750 billion of US revenue flowing through AI powered search by 2028, which is a projection rather than a measurement, but the projections and the measurements are now pointing the same way.
Put any of it next to your own organic sessions chart and the strategic question changes shape. The question is no longer how to win the clicks back. It is which of your numbers still measures demand, and most executives have not been given a straight answer to that one.
What moved, and what it costs you to ignore it
None of those rows is a marketing tactic. They are management changes, which is why they tend to get delegated one or two levels below where they belong.
Most of your readers are now machines
Cloudflare sits in front of about a fifth of the web and counts what reaches it. In June its own measurement crossed a line its chief executive had publicly forecast for late 2027.
Automated requests passed human ones for the first time, at roughly 57% of HTML requests to web content against 43% from people.
Share of HTML requests to web content, Cloudflare network, June 2026.
Machines reading your content is not new. What changed is the terms. For two decades the arrangement was reciprocal enough: a crawler took your page and an engine sent you a reader in exchange.
Crawl to referral ratios for leading AI bots have been observed anywhere from 118:1 to nearly 50,000:1, against the roughly balanced ratios of traditional search crawlers.
The spread is wide because it covers different bots in different months, and the worst offenders are training crawlers with no referral mechanism at all. Treat it as a direction rather than a precise figure. If you are a publisher this is your revenue model coming apart. If you are a brand it is something less dramatic and nearly as awkward: your content is now being read mostly by something that will not visit, will not convert, and will not appear in any report you currently run. Crawl activity and traffic have come apart. Any dashboard still reading the first as a leading indicator of the second is describing a relationship that stopped holding.
There is also the question of what those machines find when they arrive.
Across the US retail sector, average AI visibility scores came in at 75% for homepages, 74% for category pages, and 66% for individual product pages. A score of 50% means half the content on the page cannot be read by a model.
The page carrying price, availability and specification is the least readable one in the estate.
So the product page, the one carrying price and availability and specification, is the least readable page in the estate. That is precisely the page a shopping assistant is trying to read. We see the same shape in non-retail audits: the marketing pages score fine and the pages holding the actual facts score badly, because those were built for a rendering engine rather than a reader. It is a supply problem, and a bigger content calendar does not touch it.
The shortlist closes before you know it opened
Which brings the argument to the part that applies whether or not you sell anything online. If you sell a considered purchase rather than a pair of running shoes, the number that should hold your attention is not any of the ones above. It comes from the B2B side, and it predates the AI search argument entirely.
94% of buying groups ranked their preferred vendors before first contact with a seller, and bought from that preliminary favorite 77% of the time. The balance between independent research and seller engagement moved from 70/30 to 60/40.
Sixty percent of the journey happens before anyone at your company knows the deal exists.
Sixty percent of the journey happens before anyone at your company knows the deal exists. More of that sixty percent every quarter runs through an assistant that reads the reviews, lines the vendors up and hands over a shortlist. Miss that answer and you are missing from a ranking that already calls the winner better than three times in four.
Sales cannot recover this later, for the simple reason that sales has not been invited yet. The same research found something worth reading twice by anyone selling an AI enabled product: 89% of purchases included AI features, and 58% of buyers reached out to sellers early specifically because vendors had not explained those features clearly enough anywhere a buyer could find them.
That is a content gap with a revenue number attached, and one of the few in this piece you can close without anyone else's cooperation.
AEO and GEO are two budgets, not two words
None of that is fixed by renaming the SEO budget, which is roughly what the industry has done. Two labels have stuck. They are useful, but only if you read them as separate spend categories.
Answer engine optimization is work on property you own: structuring pages so a machine can lift a clean answer out of them, and making sure product data resolves without a browser. It is engineering, it is mostly a fixed cost with maintenance after it, and it sits inside your control.
Generative engine optimization is work on property you do not own. It is earning a place in the source base a model reads before it answers. The reason these cannot share a line item sits in one ratio.
A brand's own sites make up only 5 to 10 percent of the sources AI search references. In categories such as consumer packaged goods and financial services, more than 65 percent of sources are publishers, user generated content, and affiliate sites.
On that estimate, nine tenths or so of what shapes the answer about you is written by somebody else. No amount of work on your own pages reaches it.
You can hold the featured snippet and still be missing from a ChatGPT answer, which is why a single combined budget hides the failure of one half behind the success of the other. The mechanics underneath that difference, including why a citation and a mention are not the same event, are the subject of our earlier piece,How AI citations work, and why they matter.
The money moved. The capability did not.
Budget is not the constraint here. That makes this a harder internal conversation than most AI proposals, because you cannot fix it by asking for more.
CMOs allocate an average of 15.3% of marketing budgets to AI initiatives, and 70% say becoming an AI leader is critical for 2026. Only 30% report mature or fully developed AI readiness.
The capability is arriving faster than the discipline around it.
Generative engine optimization is now in use at four in ten companies, a capability that did not appear in earlier editions of the survey.
Four in ten doing the work, three in ten ready to run it. Buy a capability faster than the organization can absorb it and you get spend without a control loop, which is the same pattern the last two years of AI pilots already ran through once. You can see where it shows up.
16% of brands systematically track how they perform in AI search.
That last sample is about thirty people, so treat it as a signal and not a measurement. It stays on the page because it points the same direction as everything above it, and a weak sample shown as a weak sample beats a strong claim built on one.
Four decisions, in this order
Fix the number the board sees.
Sessions stopped being a demand metric the moment answers began arriving without clicks. Report conversion and revenue per visit by source, with AI referrals broken out from direct and organic. If your analytics still buckets assistant traffic as direct, that misattribution is costing you the argument before it starts.
Take the baseline before you fund anything.
Run the questions your buyers actually ask, on the assistants they actually use, and record three outcomes: cited, named, neither. The build is not where the money goes. The money goes into re-running it on a fixed rhythm, and skip that and what you own is a screenshot rather than a baseline. Budget for it the way you budget a brand tracker, not a project. This is the Bearing phase of ourAZIMUTH method doing unglamorous work.
Split the budget and name two owners. AEO belongs with whoever owns the site and the product data. GEO belongs with whoever owns earned media. Give both to one person and you will get whichever half that person already knows how to do, delivered competently, while the other half simply does not happen.
Write the kill criterion before the first invoice, not after the first disappointing quarter. Ours for this kind of engagement reads about like this: if a quarter of visibility work has not moved cited and named share on the prompts that matter, we stop, we put the reason in writing, and the budget goes somewhere with evidence behind it. Deciding the stop in advance is what separates a channel from a dashboard nobody is allowed to question.
Every figure above carries its date because most of them have a shelf life measured in months. Retrieval and citation behavior gets revised without announcement, and vendor reported data revises too. We check this piece on the first of the month and mark what changed. Corrections tohello@bearingbridge.com.