Every Commercial Website Will Need an Expert for Agents
Why every commercial website in the AI agent era needs a component that answers the agent directly, what three days of controlled tests showed, and what they have not yet shown.
Today’s article, Why Blocking AI Agents Is Bad Business, argued that a blanket block leaves the agent’s decision intact and removes the business from it. That raises the next question for every board that decides to let agents in: what should they find when they arrive?
Our answer is a component that most websites do not yet have. It is an expert the agent can consult, which answers only with the business’s approved and sourced facts, says plainly when nothing fits, and records what it said so the business can see the sale it won. We call ours AIVO Assist. In controlled tests run between 24 and 26 October 2026, reported in a new AIVO Standard working paper, agents with access to it chose the right business 93% of the time on realistic websites, against 47% without it.
This article sets out why we think an agent-facing expert will become as standard on a commercial website as the sitemap or the padlock, and the rules such a component has to follow to deserve that place. A disclosure belongs at the top: AIVO, Inc. builds and sells AIVO Assist, and every result below comes from our own tests. The final section states what they do not prove.
The page was written for a person. The agent arrives with a brief.
A website is built for someone who browses. It leads with what most visitors want, spreads detail across pages and leaves the rest to a phone call or a booking flow. An agent arrives with something narrower: a brief. “A lodge near the beach, hot tub, dog welcome, under £900 for the week.” “A bank that will open an account for a startup billing in three currencies.” It reads a slice of the page, tries to match the brief, and moves on.
Our research since July 2025 shows where that goes wrong. Across 27,252 AI buying conversations (104,140 turns), the decisive moment comes late, in the follow-up questions where the AI compares, checks claims and drops options. Of the brands named at the start of a buying conversation, only 12.7% were still the pick at the decision (n = 1,427, the Linkage Gap). In our banking study of 2,160 conversations, reported by American Banker in September 2026, the AI recommended someone else in 72% of conversations that opened on the customer’s own bank (502 of 702).
Independent measurements point the same way. ChatGPT’s search mostly works from a short snippet at the top of a page. Google’s grounding takes a median of about 377 words from each page. Browsing agents in one controlled study never found an offer placed a single screen below the fold. A business is judged on the slice of its information an agent happens to read.
The common cause is that the deciding fact was missing, buried or unverifiable when the agent looked for it. Prices sit behind date pickers, policies sit in PDFs, and combinations of features exist nowhere as a single sentence. No page can state every combination a brief might ask for. Something has to answer the question the page did not anticipate.
What AIVO Assist is
AIVO Assist is a business’s own expert, offered to the AI agents that visit it. The agent passes on what its customer needs, in plain words and without personal details. AIVO Assist returns one of three verdicts: fits, fits if (with the condition stated), or no suitable option. It then answers with the deciding facts, each carrying its source and the type of source, such as an independent test, a manufacturer’s specification or the business’s own claim.
Agents reach it through the channels they already use: a plainly labelled form on the page, a tool declared to browser agents through WebMCP, an API and protocol connections, and answers published in advance for assistants that read pages without interacting. The same approved knowledge feeds every channel, so every channel says the same thing. It reads and answers, and does not log in, take payment or act on anyone’s account, which keeps it on the reading side of the line today’s article drew.
Each answer carries a reference that follows the customer into the booking or purchase, and both are reported from the server into the analytics the business already uses. The business sees which answers led to which sales, and what customers asked for that it could not supply.
The working paper’s central argument is about what such an expert must know. A catalogue with a chat window answers what a product is. An agent deciding for a customer needs to know whether it fits this customer, why it beats the alternatives the agent is weighing, what the agent currently believes that is wrong, and what evidence will satisfy it. Those are properties of the decision, and they can only be found by measuring how AI systems actually decide.
That is why AIVO Assist is built from AIVO, Inc.’s measurement as well as the business’s own data. AI Win-Rate shows where and to whom a business loses the decision. The Linkage Gap shows facts the models hold but do not use. Legibility testing shows which facts agents cannot find or verify, brand.context sets the evidence each claim needs, and AI reputation measurement shows the limits the models attach to a business. Across 222 brands in our research, first-prompt visibility explains only about a third of which brands win, so most of the decision happens later, where the expert has to be ready.
What the tests showed
We ran 1,512 agent sessions over three days, using our own test agents built on Claude, Gemini and Perplexity models and reading pages the way consumer agents such as Meta’s Muse and ChatGPT’s browser do. The pass rule for every test was set before it ran. The tests climbed in levels, each asking a harder question.
The first level asked whether an agent would use the expert when told it existed, and trust what it heard. Agents reached the right decision in 119 of 119 decided sessions where they consulted it. When the business could not help, such as a hotel query reaching a holiday-park expert, the current version said “No suitable option” in 57 of 57 cases. The first version did not pass cleanly: running on a smaller model, it hedged instead of refusing in 6 of 60 cases, so the expert was strengthened and re-tested. That slip is the most useful finding of the three days, because it shows honesty has to be engineered and tested rather than assumed.
The second level asked whether agents would find and use it unprompted. On a clean, simple page about half did. On realistic pages, where the deciding facts were hidden behind booking pages or vague wording, 89% did. Claude and Gemini models used it every time; Perplexity less often.
The third level asked whether it changes the decision. The test used a controlled market of three fictional businesses, with the deciding facts behind a booking engine or in vague copy. Right recommendations rose from 47% (79 of 167) without AIVO Assist to 93% (157 of 168) with it. The pass rule, agreed in advance, was an improvement of at least 20 points with non-overlapping uncertainty ranges. The result was 46 points, with ranges of 40–55% and 89–96%. On a clean page that already stated everything, it made no difference, which is the result we would want an honest component to produce. In 149 of 168 recommendations the agent used a fact that appeared only in AIVO Assist’s answer, such as a lodge with a hot tub, dogs welcome, £860, three minutes from the beach. Without it, that fact never appeared.
The final test asked whether the business could see what happened. Plausible’s standard tracking recorded none of the agents’ 314 page visits, because it is designed to ignore automated browsers. AIVO Assist’s own events arrived in full: 79 bookings, 31 credited to AIVO Assist, none credited wrongly, matching our log exactly after one attribution fault was found and fixed.
The limits are as important as the results. These were our agents on our test sites. We have not yet shown that consumer agents such as Muse, ChatGPT’s browser or Dots behave the same way, that the result holds on a live business website, or that it produces real sales from real customers. Those are the next tests, and we will publish them whatever they show.
Why it becomes standard
The web already has a list of components that every commercial site carries because machines visit it. The robots.txt file told crawlers where not to go. The XML sitemap, backed by Google, Yahoo and Microsoft from 2006, told them what existed. Schema.org markup, launched in 2011, told them what a page meant. HTTPS was optional for most sites until search engines rewarded it and browsers began labelling plain pages “Not secure”; within a few years a site without the padlock looked broken.
Each became standard under three conditions. The visiting machine gained something from it, so it looked for it. The cost to the site was small and the cost of leaving it out grew. And there was a shared form, so a site could adopt it once and serve every visitor.
An agent-facing expert already meets the first condition. Agents in our tests used AIVO Assist unprompted 89% of the time when the page was hard to read, most likely because asking was quicker than searching the page. It meets the second: it is a component added to an existing site, and the cost of leaving it out is the shortlist the business never learns it missed. The third condition is the one still open, and it is where this article’s argument leads.
The agent protocols are still settling. WebMCP, the Personal Agent Protocol, Web Bot Auth and the commerce protocols from OpenAI, Google and the payment networks will not all survive in their current form. An expert that answers in plain language, on the page, works with whichever of them wins, and can add each as a further door to the same set of approved facts. A business that waits for the protocol war to end before deciding what its agents should be told will have waited too long.
The rules a canonical component must follow
A component earns a standard place only if visitors can trust every instance of it. Agents are already trained to treat text on websites with suspicion, and a sales bot dressed as an expert would teach them to ignore all of them. The working paper sets out six principles AIVO Assist follows. We intend to publish them, with the two tests below, as an open specification under AIVO Standard, so that any vendor, or any business building its own, can be held to the same standard.
It says only approved, sourced statements, and cannot introduce a claim of its own. It gives an honest “no suitable option” whenever nothing fits, without hedging or redirecting. It gives the agent no instructions: it describes itself and answers questions, and never tells the agent what to do or what to recommend.
It keeps parity, so agents are shown the same facts as people and never a different business. It needs no personal data, working from the need (“a family of four with a dog”) and never the person. And evidence outranks assertion: where independent evidence contradicts the business’s claim, the expert does not state the claim as fact, does not dispute the evidence, and reports the conflict to the business to fix at source.
A standard also needs two tests. Its effect must be measurable by anyone, with and without the expert, against a rule agreed before the measurement and judged on the agent’s final choice. And each answer must be traceable to the sale it led to, in analytics the business already trusts.
Two governance rules complete it. What agents ask stays each business’s own, and one business’s demand is never shown to another. And AIVO Assist is not deployed for political candidates, parties, elected officials or advocacy positions, which AIVO, Inc. measures but never remediates. A component that cannot meet these rules has no claim to be standard, whoever makes it.
What boards should do now
Start with the facts that decide your sale. For your ten most valuable customer briefs, check whether an agent can reach the deciding fact in one step from the page it lands on, and record how often it cannot. That gap is the size of the opportunity.
Then give someone ownership of what agents are told. Today it usually falls between marketing, which owns the message, and technology, which owns the site, so the answers belong to nobody. Decide which facts are approved, who keeps them current and who signs off on the answer “No suitable option”.
Finally, ask anyone who offers you an agent-facing expert, AIVO, Inc. included, for the measured version: the with and without test, the pass rules set in advance and the result on the agent’s final choice. If the answer is a dashboard of mentions, keep looking.
Letting agents in keeps a business on the shortlist, and answering them well is what wins the decision. We expect the expert that does the answering to be as unremarkable on a commercial website in a few years as the padlock is today, and the evidence so far points that way. The tests that would settle it, with consumer agents on live websites and real customers, come next.
Disclosure and sources
AIVO, Inc. publishes AIVO Journal and builds and sells AIVO Assist, AIVO Meridian and AIVO Win Rate. The AIVO Assist results are from controlled tests with AIVO, Inc.’s own agents on test websites, 24 to 26 October 2026. They are DOI-anchored and published for scrutiny, not peer reviewed.
Beyond Visibility: the Linkage Gap, AIVO Standard (DOI 10.5281/zenodo.20848721)
AI Win-Rate: A Board Metric for AI-Mediated Buying (Methodology v2.0)
The Product Data Legibility Gap, AIVO Meridian (DOI 10.5281/zenodo.20322459)
RESONEO, What ChatGPT pulls, what it shows, what it cites
DEJAN, How big are Google’s grounding chunks?
Nitu, Mühle and Stöckl, Machine-Readable Ads (arXiv:2507.12844)
Why Blocking AI Agents Is Bad Business, AIVO Journal
Comments ()