AI agents now do the first round of shopping for your customers. They read everything you’ve published, believe all of it, and repeat it with confidence. We measured what they say.
By Jason Turner, Co-Founder & CEO, IDLab · Last updated · Data: 84 hand-scored queries, August 2026
The short version
AI agents don’t misread your brand. They misquote your facts. Accuracy is yours to control. Citation is the engine’s whim. Win the fight you can win.
The most literal customer you’ll ever have
As marketers we’ve spent our careers honing our skills to persuade people who aren’t paying attention. We have jingles for the distracted, billboards for the bored and loyalty points for the forgetful. Our whole craft is built on a happy truth: humans are gloriously irrational, and the irrational bits are where the money is.
Now meet a customer with no irrational bits.
It doesn’t get hungry at 4 p.m. It can’t be charmed by a nice serif. It’s never seen your Super Bowl ad and wouldn’t care if it had. It reads everything you’ve ever published, believes all of it, and repeats it to the customer you’re really trying to get to with the calm confidence of a person who’s never once been wrong at a dinner party.
Including the PDF from 2023 you forgot was still online.
That customer is an AI agent. Marketing to it now has a name: B2A, business to agent. Most of what’s written about it is either breathless or dismissive. I wanted to know what the machine actually says, so I measured it. The results are funnier and more alarming than the panels suggest.
What B2A actually is
B2C sells to people. B2B sells to companies. B2A sells to the software that people and companies send shopping for them.
Your customer asks ChatGPT, Perplexity, Gemini or Claude which insurer covers what, which flight is cheapest, which software tier includes SSO. The agent searches, compares, picks a winner and writes one answer. Increasingly it also books, buys or signs up. The customer reads one answer. There’s no page two. Often there’s no page one either.
Expedia’s CMO, Jochen Koedijk, put it neatly in May: “Agents are becoming a new audience, in addition to businesses and consumers.” The key words are in addition to. Your human customers haven’t gone anywhere. They’ve hired a personal assistant, and that assistant now runs the first round of interviews. BTW, you’re not invited.
How it works: four steps, zero sentiment
An agent does roughly what a very fast, very literal intern would do with your website:
- FindIt searches or crawls for sources. If your site blocks it, hides the facts behind JavaScript or bakes them into an image, you aren’t in the running. You barely exist, if at all.
- ExtractIt lifts the specific fact: price, eligibility, deadline, availability. A table it can read beats a paragraph it has to interpret, no matter how beautifully written your paragraph is.
- ReconcileIt checks what it found against other sources. When your site, your reseller and a review blog disagree, it picks one. It doesn’t always pick you (unless you’re on top of your B2A game).
- Answer, then actIt writes the answer, names a source and, more and more often, takes the next step for the customer, which could include transacting on the customer’s behalf.
Notice what’s missing from that list: your brand story, your tone of voice, your hero image and the agency fees that went into all three. The agent skips them the way you skip convenience store hotdogs.
You could say that agents are inherently lazy. They forage for the path of least resistance to the fact. Your job is to make that path lead to you with the least amount of friction.
Is B2A a fad?
The hype is early. The plumbing is not.
The case for “calm down”: Expedia is building a whole marketing function aimed at agents, yet in May it said less than 1.5% of its traffic came from answer-engine optimization. As of July, no major retailer had publicly reported more than 5% of online sales from purchases an agent completed on its own. If you’re being told agents already run your funnel, you’re being sold hyperbolic BS. But wait a minute. That doesn’t mean Expedia and others are wrong. They’re just out of the gate early, which could be hugely advantageous to them. For the record, I think it will be.
The case for “pay attention”: Between April 2025 and January 2026, Mastercard, OpenAI, Visa and Google launched five competing protocols that let agents browse, authenticate and pay. In June, Visa and OpenAI announced they would build agent-led payments into OpenAI’s products. Salesforce estimates AI and agents influenced about 20% of global online holiday sales last year, roughly $262 billion.
Visa doesn’t build payment rails for a fad. Fads don’t get millions of dollars of investment and fraud teams from a company whose entire business model could be disrupted. And zoom out for a second. Wall Street and tech giants have built a modern digital bypass to handle automated, instantaneous institutional finance, and the major card companies are building new protocols to act as the financial rails for AI shopping agents and autonomous software. So what? The money itself is being re-plumbed for software that buys things. When the payment is instant and the buyer is a machine, nobody pauses to double-check the price. Whatever the agent believes about your offer is what gets transacted.
And the twist nobody wants to hear. The fashionable line is that agents ignore brands. Skift wrote that brand recognition “means nothing to AI agents.” I think that’s half right, which is the most dangerous kind of right.
Brand doesn’t die in B2A. It moves upstream. A human uses brand as a shortcut at the shelf. An agent uses something similar earlier: which sources it trusts, which facts the wider web agrees on, which name it attaches to the question. Reputation becomes consensus. That’s still brand. It’s just brand with a fact-checker standing behind it.
So what does that mean for you? When your buyer brings a fact-checker, some old habits stop working:
- Claims you can’t back become liabilities. “Industry-leading.” “Best-in-class.” “#1 rated.” An agent goes looking for the proof. If it can’t find any, the claim gets ignored. If it finds the opposite, it quotes that instead.
- Spin stops converting. Clever framing works on a hurried human. An agent reads the terms, does the math and reports the real price, fees from paragraph nine included.
- Sloppy content gets skipped. Facts trapped in images, offers scattered across microsites, pages that contradict each other. The agent won’t untangle your mess. It’ll take the answer from someone tidier.
- Your partners’ words become your words. Resellers, affiliates, old press releases and forgotten landing pages all count as evidence about you. Their stale claims are now your problem.
- Reviews become your spec sheet. If you say the battery lasts ten hours and a thousand customers say four, the agent may well side with the customers.
- The gap between pitch and product gets exposed. When the product doesn’t match the marketing, agents notice on every query, at scale, forever.
- Consistency becomes authority. The brand that says the same true thing everywhere earns the agent’s confidence. Repetition used to build memory. Now it builds proof.
What I found when I checked
A hunch is a fine place to start and a terrible place to stop. So my team and I went and measured it.
We ran 84 hand-scored queries across ChatGPT, Perplexity, Gemini and Claude, asking about three organizations’ published offers: what’s the deal, who qualifies, how do I get it, is it worth it. Each query ran three to four times. We scored only the answers that held steady, against each organization’s actual published terms. Every capture was timestamped, in August 2026.
The failures showed up across several verticals; basically anywhere a business has regional pricing, partner channels or offers that expire. That covers retail, travel, insurance, telecom, energy, franchises, healthcare and most of B2B software.
Being quoted isn’t the same as being right
VIEW AS TABLE
| Organization | Own site cited | Accurate |
|---|---|---|
| Single-region organization | 100% | 100% |
| Large multi-region organization | 79% | 93% |
| Two-region organization | 96% | 61% |
Five things stood out.
01The machine isn’t lying about you. It’s quoting you.
The organization with one territory and clean, current web pages was accurate 100% of the time, on every engine. The one with two regions and a drawer full of aging PDFs was wrong or incomplete 39% of the time. And here’s the punchline: the engines cited that company’s own website 96% of the time. On one question, all four engines cited the company’s own site and none got the answer fully right. The source was the company. The source was wrong.
02Your least-loved document is your most-quoted source.
Engines love PDFs, and for a sensible reason. Your PDF states the number in a tidy table, while your web page wraps it in three paragraphs of aspirational copy. So the document you update least is the one the machine trusts most. A stale PDF is just confident misinformation with your logo on it.
03Expired offers live forever on other people’s websites.
In one durable goods scenario, two engines repeatedly told customers they could claim a $2,000 incentive that had expired seven months earlier. The usual source wasn’t the company. It was resellers’ sales pages that nobody had updated. In one answer, 11 of 14 citations pointed to a single reseller.
04Being right doesn’t mean getting the credit.
One organization was 93% accurate and still won the citation only 54% of the time. Aggregators and partners took the credit for its facts. Across all 84 answers, the company’s own site appeared 92% of the time but was the main source only 65% of the time. A third of the time, on a question about a company’s own offer, somebody else was the authority.
05The engine matters more than you do.
Same questions, same companies. The company was the cited authority 88% of the time on ChatGPT and Perplexity, and 43% of the time on Claude and Gemini. You don’t choose your customer’s engine. Treat citation like weather: plan for it, but don’t think you can control it.
Who gets the credit depends on the engine
VIEW AS TABLE
| Organization | ChatGPT + Perplexity | Claude + Gemini |
|---|---|---|
| Single-region organization | 93% | 57% |
| Large multi-region organization | 86% | 21% |
| Two-region organization | 86% | 50% |
| All three | 88% | 43% |
Accuracy is yours. Citation is the engine’s whim. Win the fight you can win.
How to optimize for B2A, step by step
None of this needs new technology or a new agency. It needs your facts made legible to a machine. This is the order I’d do it in.
- Ask the machine first. Write down the 10 to 20 questions customers actually ask: price, availability, how/where to buy, eligibility, unique selling points, use/application, etc. Run each on four engines, three times. Score two things separately: was it right, and who got cited? That’s your baseline, and it takes about a day. Most companies have never done it. Most of you are about to be surprised.
- Give every fact one home. Pick a single canonical page for each offer. If the same price lives on a landing page, a promo microsite, a partner portal and a PDF, the machine sees four witnesses telling four stories. It believes none of them completely. So what? A confused agent hedges, averages or goes with whichever version looks most confident, which is often your partner’s expired promo. Then your customer shows up expecting a price you don’t offer, and your service team and your margin pay for it. One home means one answer, and one answer is the one the agent repeats.
- Lead with the answer. The number goes in the first sentence, in plain text: “Up to $500 off, through December 31, for customers in X.” Save the brand poetry for after. The machine has already left. I know this goes against much of our traditional training of “emotional connection” but times change. I’m not saying that’s gone or not important; I’m saying the order of operations has changed.
- Get facts out of PDFs and pictures. Every rate card, pricing table and terms sheet becomes real HTML text and tables. Keep the PDF for the people who love PDFs. Both of them.
- Tag every fact by where and when. Region, segment, start date, end date. Untagged facts get blended, and a blended answer fits no real customer anywhere. For a retailer, that means online-only versus in-store, member versus non-member price, which stores actually stock it, and the hour the sale ends. Skip it, and the agent promises your shopper a $49 price from an online-only deal that ended Sunday. Then a store associate gets to explain why.
- Date it and give it an owner. A visible “last updated” date on your site or digital asset, a machine-readable
dateModified, and a named human whose job is keeping it true. Freshness without an owner is just a timestamp. - Speak machine. Add structured data (Product, Offer and FAQ schema; yes, this is also just good SEO/AEO practice) and keep product feeds current. Check that your robots.txt and bot protection aren’t blocking the AI crawlers you actually want. Pilot an
llms.txtfile. It’s cheap and still unproven, so treat it as an experiment, not a strategy. We’ve been doing it for a while, and we think it’s helping a little, but it’s definitely not hurting anything, and it costs nothing to do. - Arm your resellers. Partners, dealers, affiliates and review sites will repeat your facts whether you help them or not. Give them a clean, current, region-tagged feed and ask them to link back. Turn your interceptors into corroborators.
- Re-run it every month. Engines change weekly. Track accuracy and citation share over time, not impressions. This is gardening, not construction. Nobody plants a garden once.
Pitfalls, and how to dodge them
- Measuring the wrong thing. Citation share is a vanity metric you don’t control. Accuracy is a sanity metric you do. Report both, but manage to accuracy.
- Locking the door on paying customers. In September, Resy deactivated a diner’s account and canceled his reservations. His offense? He sent his personal AI agent to find him a table. The agent did what agents do: it checked constantly and moved fast when tables were released. Resy’s systems couldn’t tell a customer’s own assistant from a scalper’s bot, so they treated a paying customer like a threat. Two days later, they reinstated him. Think about what that means. He did nothing wrong. He delegated. Most bot and fraud rules were written for a world where any non-human visitor was up to no good. That world is ending. If your defenses can’t tell the difference, you’ll punish your most engaged customers for being early adopters, and the booking goes to a competitor whose door is open. So give good agents a front door: a proper agent endpoint or API, rate limits designed around how agents actually behave, and a way to verify who sent them. Visa’s Trusted Agent Protocol exists for exactly that.
- Keyword stuffing, the sequel. Don’t write pages for robots that read like ransom notes to humans. What the machine needs (findable, extractable, specific, current, corroborated, region-clear) is exactly what a busy person needs too. If the page got worse for people, you did it wrong.
- Mistaking a protocol for a strategy. There are five agent-commerce protocols and no winner yet. Build clean data any of them can use, and don’t bet the business on one vendor’s spec.
- Buying snake oil. Anyone who promises to “guarantee your brand appears in ChatGPT” is selling you weather insurance. Ask how they measure accuracy, and watch them blink.
- The dead end. If the agent cites your PDF, your customer lands in your PDF. No button, no next step, no sale. Every source the machine might cite needs a path to buy.
What’s coming in the next 12 months
Here are the dates and numbers, with sources so you can check my homework. Forecasts are forecasts, and I’ve labeled my own.
- November–December 2026Adobe forecasts that traffic from AI chat and browser services to US retail sites will rise 130% year over year, on a record $275.1 billion in online holiday spend. That’s clicks, not agent purchases, but it’s the on-ramp.
- Through 2027The payment rails go live in pieces. Visa and OpenAI announced their integration in June without a launch date. My call: the five protocols won’t produce a winner in the next year, so merchants will have to support several.
- Any quarter nowWatch for the first major retailer to report more than 5% of online sales from purchases agents completed on their own. As of July, nobody had cleared that bar. When someone does, every board will ask its CMO about B2A the following Monday (except Expedia’s).
- By 2028Gartner predicts AI agents will intermediate more than $15 trillion in B2B spending. If you sell to businesses, the agents are coming for your procurement process first.
- By 2030McKinsey estimates agentic commerce could orchestrate up to $1 trillion of US retail revenue.
The number I’m watching most comes from my own audit. Only 1 of 84 answers sent the customer to a third party to actually transact. Today a wrong answer costs you a click. Once agents complete the purchase, the source the agent trusts becomes the store the agent buys from.
My bet: within 12 months, that 1-in-84 becomes the most expensive statistic in your marketing plan. The companies fixing their facts now are buying a head start that’s hard to make up later.
Your first audience is a machine
For a hundred years, we’ve had one job in marketing: get a human to take the action you want them to take. B2A adds a second one: get a machine to believe you.
The good news is that the machine is the easiest customer you’ll ever persuade. It doesn’t want a jingle, a discount or a celebrity. It wants the truth, stated plainly, somewhere it can find it, with a date that proves it’s fresh.
The bad news is that it already believes everything you’ve ever published. It has read your fine print. All of it. Including the bits you forgot.
Sources
- Audit data: IDLab, 84 hand-scored captures across four AI engines, August 2026.
- Skift: B2A? AI Agents Are Travel’s New Audience (Expedia quote, 1.5% figure, brand recognition line)
- State of AI: agentic commerce prediction (5% retailer bar; Salesforce 20% / $262B influence figure)
- Elogic: ChatGPT Commerce & Agentic Shopping Statistics 2026 (five protocols and launch dates)
- Digital Commerce 360: Visa, OpenAI work together to support agent-led payments (June 2026; McKinsey $1T estimate)
- Relevant Audience: Adobe 2026 holiday forecast ($275.1B; AI traffic +130%)
- Digital Commerce 360: Gartner, AI agents and $15 trillion in B2B purchases by 2028
- Lorikeet: What Is B2A? (Resy account deactivation, September 2026)