The Agent That Doesn't Knock Twice

Personal agents like Muse decide in a single pass. What that changes about winning the sale.

Postmen ring twice. The AI agent shopping for your customer rarely comes back with a question at all. In research published by AirOps this week, Meta’s Muse asked its user a question in 0.2% of answers, against 39.5% to 91.5% for ChatGPT, Gemini and Google’s AI Mode.

That figure does not tell us how many internal steps Muse takes before it answers. It tells brands something commercially important: they cannot assume there will be another conversational turn in which to supply missing evidence or correct a misunderstanding.

Much of what AIVO, Inc. has published argues that AI buying decisions are made late, in the follow-up questions where assistants compare options, check claims and drop businesses. For chat assistants that remains true. For agents that rarely hand the conversation back, a brand cannot count on a later turn. In this article, a one-pass agent means exactly that: one that makes its recommendation without asking its user for further input.

This article sets out what changes, what our own tests suggest about why agents drop a business when evidence is missing, and how brands should measure different kinds of AI buyer.

Two kinds of AI buyer

AirOps compared answers from Muse, ChatGPT, Gemini and AI Mode between 23 and 27 October 2026. The differences are large enough to describe two different kinds of buyer.

Source: AirOps Research, October 2026.

Chat assistants read a handful of sources, write a long, structured answer and hand the conversation back to the person. Muse reads around four times as many sources, from four times as many domains, and returns a few sentences. AirOps’ own reading is that Muse is oriented toward reaching a decision rather than carrying on a conversation. That fits the direction personal agents are taking: an agent that shortlists, books or buys on someone’s behalf is valued precisely for not coming back with questions.

Where brands lose, and what changes

AIVO, Inc.’s research has measured buying conversations turn by turn since July 2025. Across 222 brands in our AI Win-Rate measurement, how visible a brand was in the first answer explained only about a third of which brands won, as reported in our AIVO Assist working paper. In a separate study of 1,427 brand probes, only 12.7% of the brands named at the start of a buying conversation were still the pick at the decision (the Linkage Gap). The models know the brand; they drop it in the follow-up turns, where they compare, check and narrow.

Those findings come from conversations with chat assistants, where the later turns exist. With an agent that rarely hands the conversation back, whatever comparing, checking and narrowing it does happens out of the customer’s sight, before it answers, and Muse averages roughly 25 citations per answer. The facts a brand might once have supplied in the third turn of a conversation need to be available when the agent forms its answer.

This also exposes the limit of measuring mentions. Chat assistants decide later than the first prompt, so a first-prompt mention says too little about them. One-pass agents decide in the first answer, but on far more evidence than a mention, so a mention says too little about them as well. In both cases the measure that matters is the decision itself.

One pass punishes what cannot be checked

Our tests with AI agents point to two gates where a business is dropped. At the first, the agent cannot find the deciding fact. At the second, it finds a claim but cannot reach or verify it. In an exploratory run of 135 agent decisions about three UK holiday parks (published here), 18 of the 19 cases in which an agent cut a park over evidence concerned price, fees or availability for the customer’s dates. The facts existed, but not in a form the agent could confirm.

A baseline we ran this week for a consumer brand, with agents on ChatGPT, Gemini and Perplexity searching and browsing the live web, shows the second gate in the agents’ own reasoning. When the brand was found but lost, agents explained that it cited an independent study but gave no link on the product page that would let them verify it, and that the only evidence they could locate was the brand’s own summary. The study existed and was published. The product page did not link to it.

In a conversation, a person might ask where the study is. The practical consequence of a 0.2% question-back rate is that brands have little reason to expect that second opportunity. An agent may discount the claim, qualify its recommendation or choose an alternative, and how often each happens is an empirical question. The risk is that a brand loses before it has another chance to make its case.

The reverse also holds, at least under test conditions. In a controlled market of three fictional businesses, with the deciding facts behind a booking engine or in vague copy, agents recommended the business that met the customer’s brief in 79 of 167 sessions without an expert on the page (47%) and in 157 of 168 with one (93%). The same sites, briefs and agents were used in both conditions, and the pass rule was agreed before the test.

The deciding fact appeared nowhere on the websites. It was in 149 of 168 recommendations made with the expert, and in none made without it. These were AIVO, Inc.’s own agents on test sites; a test on a live business website, with real customers, is next, and the full design is set out in the working paper.

The brand’s own site is one voice among many

Muse draws 17.6% of its citations from review and user-generated sites, five times ChatGPT’s share, and averages roughly 25 citations per answer. Those figures do not show how much weight each source carries in the final choice. They do show the risk of relying on owned content alone.

That changes what counts as evidence. A claim on the brand’s site that is confirmed by reviews, a retailer listing and an independent test carries weight. The same claim standing alone, or contradicted elsewhere, is likely to be discounted. Our working paper on agent-facing experts makes the same point as a design rule: where independent evidence contradicts a business’s claim, the claim should not be stated as fact. One-pass agents apply that rule whether brands design for it or not.

What brands should do

Make every deciding fact reachable and checkable from wherever an agent lands. If the decision turns on a price, a policy or a study, the agent should find it, and its source, without passing through a booking engine, a PDF or a support page. For the brand above, the immediate fix for the gap agents identified is to link its published evidence from the product page.

Make the facts agree wherever agents read them. A one-pass agent compares the brand’s site with retailer listings and reviews in the same read. A price that differs, or a claim that appears only on the brand’s own page, is a reason to move on.

Make the evidence available during the agent’s existing visit. The first priority is to publish the facts and sources an agent needs to verify a claim. Asking the customer a question and using a tool on a page are different operations, and an agent that rarely does the first may still do the second while it browses.

Our own tests point the same way. On a clean test page, agents that skimmed only the first screen, as personal agents tend to, consulted the expert in 82% of sessions; the same agents reading the whole page did so in 56%. Research bots that cannot use forms never did, which is why the same facts also need publishing in advance.

Where an agent can query a page-level expert, through a labelled form or a tool declared to browser agents, the site can supply missing context without another question to the customer. AIVO Assist is designed for that second task; it complements accessible evidence and does not replace it. We set out the wider case in Every Commercial Website Will Need an Expert for Agents.

Measure the decision under each buyer’s actual conditions. For chat assistants, that is the recommendation at the end of a defined multi-turn buying journey, which shows whether a brand survives comparison, verification and narrowing. For browsing and shopping agents, it is the final recommendation or selection after a defined task, with the agent’s own search, tools and evidence.

The categories will blur as agents evolve. The durable approach is to classify agents by what they actually do: whether they ask the user for clarification, search externally, use tools, revisit sources and make a final selection. A single blended score would hide where a brand is losing.

Then measure the sales themselves. Standard analytics scripts are built to ignore automated browsers; in our tests, a site’s standard script recorded none of 314 agent visits. A brand that wants to know what one-pass agents are worth needs each answer, and each order that follows, reported from its own server into its own analytics.

The agent will rarely knock twice. A brand cannot assume it will ask for the evidence it needs; the evidence has to be there when the agent decides.

Disclosure and sources

AIVO, Inc. publishes AIVO Journal and builds and sells AIVO Assist, AIVO Meridian and AIVO Win Rate. The evidence in this article is of three kinds. The Muse comparison is AirOps’ research on how four AI systems answer, reported as published; AIVO, Inc. has not reproduced it, and it does not show why brands win or lose. The Linkage Gap and AI Win-Rate findings come from AIVO, Inc.’s published studies, with methods set out in the papers below.

The agent experiments are preliminary. The holiday-park run is exploratory and published at holiday.aivowinrate.com; the consumer-brand baseline (127 agent runs) is reported without naming the brand; and the expert test used AIVO, Inc.’s own agents on test sites. Our papers carry DOIs so they can be found and inspected. A DOI is a permanent identifier, not a quality guarantee, and none of this work is peer reviewed.

AirOps Research, Muse vs ChatGPT, Gemini and AI Mode, 23–27 October 2026, shared by AirOps on LinkedIn, October 2026

Sheals and de Rosen, AIVO Assist: Seller-Side Decision Experts for AI Shopping Agents, AIVO Standard, 27 October 2026 (DOI 10.5281/zenodo.23694598)

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)

AIVOCORP