The Agent Commerce Index explained: how 225 UK and US retailers score for AI agents

Direct answer: The Agent Commerce Index is a quarterly benchmark of whether AI shopping agents can find, read, trust and buy from online retailers. The Q4 2026 edition scored 225 UK and US retailers across 8 sectors, with a median score of 63 out of 100. Only 8 of the 116 stores we could basket-test (7%) let a page-fetching agent reach checkout, against 96 (83%) for an agent that browses like a person. The most common failures were missing agent protocol endpoints, and shipping and returns terms that appear only as prose, not as structured data.

Every figure in this guide comes from the live Index page or its published data file (CC BY 4.0), checked on 22 October 2026. We have not added figures that are not published there.

What the Index measures

The Index applies the open Agent Conversion Score specification to a fixed list of retailers. The score answers two questions, published in the spec as two sub-scores:

The headline score combines them as a funnel, not an average. The spec gives the formula:

Agent Conversion Score = Agent Choice × (0.7 + 0.3 × Agent Checkout / 100), then caps.

This means a closed checkout limits the score rather than being averaged away. A site agents can recommend still converts through a human handoff, so it is not zeroed. Three checks cap the score when they fail:

Grades are A (85 to 100), B (70 to 84), C (55 to 69), D (40 to 54) and F (0 to 39).

One version note: the Q4 2026 Index was scored against spec v0.5. The spec is now at v0.6 (published 17 October 2026; changelog entry 20 October 2026). The main v0.6 change is that when a store's basket and checkout cannot be observed, Agent Checkout is reported as "not measured" instead of being scored from the agent-protocol checks alone. The Index already reports unseen checkouts as "not measured" in the same way.

How the Q4 2026 edition was measured

From the method notes, measured on 19 October 2026:

• Reading the store. Each store was read as a real browser and as four agent identities: GPTBot, ChatGPT-User, PerplexityBot and ClaudeBot. Product pages were sampled from the store itself.
• The basket test. The test added one in-stock product to an anonymous basket as ChatGPT's agent and opened the checkout page. "Nothing was ever submitted."
• Two kinds of agent. A page-fetching agent is "ChatGPT's agent making plain page requests, without running a browser". A browsing agent is "the same steps in a real browser". Where the checkout turned the agent away, the steps were repeated as an ordinary browser and, on Shopify, in a real browser.
• Labels. "Browser only" marks a store whose checkout opened in a real browser but not to plain requests. "Agents refused" marks a store that let a browser through but not the agent. A checkout that refused every automated request is "unverified" and left out of the basket figures. A store that returned HTTP 429 (slow down) is also "unverified", not a failure.
• The recommendation check. ChatGPT, Claude and Perplexity, with web search on, were asked two shopper questions for the store's kind of product, without naming any brand. Because answers vary between asks, this is "reported beside the score, not inside it".
• What is not reported. The edition does not report a "recommended but not buyable" figure. For most brands the basket could not be tested, and "an Agent Checkout score below 40 is not proof that an agent cannot buy".

Coverage

Each of the 8 sectors listed 30 retailers. The 15 unscored brands are named on the Index page rather than guessed at.

Headline results

What the headline means

Agents can mostly read retailers, but rarely buy without a browser. The median Agent Choice (74) is well above the median Agent Checkout (57). The gap between the two agent types is the starkest result. Most stores that let a person-like browser through did not serve the same journey to an agent making plain requests under its own identity. The Index does not break this gap down by cause. To see where agents stall on your own purchase flow, the Simulated Agent Buyer runs browser agents through it and records every step.

Being recommended and being ready are different things. 38% of brands were named at least once when a shopper asked for their kind of product. Of the named brands we could basket-test, 34 of 39 opened checkout to a browsing agent. Recommendation depends on more than technical readiness, which is why we report it beside the score rather than inside it. Recommendation Monitoring tracks it weekly for buy-intent prompts in your own category.

Results by sector

From the Index page and data file. "Fetching" and "Browsing" are the share of basket-tested stores in that sector where each agent type reached checkout. "Named" is the share of brands named by AI.

Brand names are public observations from the Index, not clients. Three patterns are worth noting. Sector basket samples are small (9 to 19 stores), so treat them as directional.

• Beauty and fashion lead on Agent Choice (medians 81 and 78), yet fashion has the lowest AI-recommendation rate (7%). Readable sites are not automatically the ones assistants name.
• Consumer tech has the highest recommendation rate (64%) but the lowest median score (58). Our interpretation is that assistants name well-known tech brands from what they already know about them, even where the sites are harder for agents to use. The Index does not test this.
• Footwear and outdoor, and sports and fitness, have the weakest checkouts (median 50, with browsing agents reaching checkout at only 60% and 69% of tested stores).

The most common failures

The data file lists the ten checks failed by the most scored sites:

How to read this list:

• Ranks 1 and 2 are near-universal, so they barely separate stores. Protocol endpoints such as ACP and MCP are new, and they carry 2 points each in the spec. They matter where those channels are open to you, which varies by market and platform, but they are not where most stores lose ground to competitors. Our Hosted Agent Layer is built for stores that want them.
• Ranks 3, 4, 6, 8 and 10 are structured-data gaps with an established fix. Shipping details, return policies, ratings, product identifiers and sale prices (with validity dates) all appear in Google's merchant listing structured data documentation. They are the cheapest points available to most retailers, and the spec gives the fix for each check.
• Rank 5 (llms.txt) is weighted lightly in the spec (1 point) for a reason. Google's guidance on AI features says you don't need "AI text files" to appear in AI Overviews or AI Mode, and an Ahrefs study of 137,210 domains found that 97% of the llms.txt files it observed received no requests in May 2026. Add one if you like, but not before the structured-data fixes.
• Rank 7 (captchas on basket or checkout) blocks transactions directly. The spec's fix is to move anti-fraud to the payment step or exempt verified agents. Where that sits against your fraud risk is a business decision.

How to compare your store

1. Score your store. The free Agent Conversion Score applies the same published specification (now v0.6) to your homepage, product pages, cart and checkout, in about a minute.
2. Compare against your sector median, not the overall median. A 61 is middling in beauty (67) but above the median in consumer tech (58).
3. Look at the sub-scores separately. An Agent Choice in the 80s with an Agent Checkout in the 40s points to checkout friction, not content.
4. Check whether you hit a cap. A score of exactly 40, 50 or 60 can mean ACC-10, LEG-02 or TRX-01 failed. Fix a cap first, because nothing else can lift the score past it. For ACC-10, the AI access checker shows what your server returns to each agent.
5. Work down the failures table. If you fail the structured-data checks most competitors also fail, fixing them is a quick way to move ahead of your sector.

Using the data

• Licence. The data is published as JSON and CSV under CC BY 4.0. Reuse it with attribution to ConversionVector and a link to the Index.
• Corrections. Any listed brand can request a free re-scan or a correction at hello@conversionvector.com.
• Limits.
  • The Index is a snapshot of one day (19 October 2026).
  • Recommendation answers vary between asks.
  • Basket tests covered 116 of 225 stores.
  • The sample is UK and US retailers, chosen by sector. It is not a random sample of all ecommerce, and it says nothing about EU, Canadian, Australian or other stores.
  • Scores describe how agents experience a site, not its sales or commercial performance.

Common mistakes when reading benchmarks like this

• Treating the score as a ranking of how well a business is run. It measures agent readability and transactability only.
• Comparing across spec versions without checking. The Q4 2026 Index used v0.5. Your free score uses the current spec, so small differences can come from method changes.
• Over-reading small sector samples. A 100% browsing rate from 9 stores is not a law of the sector.
• Assuming a low checkout score proves agents cannot buy. The method notes say explicitly that it does not.

All figures are from the Agent Commerce Index Q4 2026 (measured 19 October 2026, spec v0.5) and the Agent Conversion Score specification v0.6, as published on conversionvector.com and checked on 22 October 2026. External sources checked on 22 October 2026.