Search gave every brand a position. You might be first or forty-first, but you were on the list, and with enough work you could climb it. An AI shopping agent does not publish a list. It reads widely, narrows quietly, and tells the shopper about a handful of products. That handful is the candidate set.
What is the candidate set?
The candidate set is the group of products or brands an agent shortlists for a specific shopper question before it writes its answer. It is what is left after the agent has filtered everything it knows, and everything it can look up, against the budget, the use and the constraints in the question.
The idea is not new. Marketers have talked about the "consideration set" for decades: the few brands a shopper keeps in mind when they are ready to buy. What is new is who builds it. The shopper used to assemble their own shortlist from search results, ads, shelves and friends. Now they hand that job to an agent and receive the shortlist ready-made, often with one name already at the top.
Three properties make the candidate set different from a search results page:
- It is small. A typical answer names three to five options. There is no page two to be found on.
- It is built per question. "Best running shoes under $150" and "best running shoes under $150 for wide feet" produce different sets. You are not in or out once. You are in or out thousands of times, one question at a time.
- It is invisible from the outside. No agent gives brands a dashboard showing which sets they made. You only find out by asking the questions yourself.
Why do rankings no longer describe what happens?
Rankings no longer describe what happens because the shopper never sees a ranked list of pages; they see a finished answer. Position seven on a results page still got some clicks. "Position seven" in an agent's reasoning gets nothing, because the answer stopped at five.
That changes the shape of the prize. In search, visibility fell away gradually as you went down the page. In an agent's answer it falls off a cliff. Either you are one of the names, or you are absent. This is why we say being considered at all is the new position one: the step from "not in the set" to "in the set" is worth more than any move within it.
It also changes what good work looks like. Classic SEO asks how to move a page up a list. Agent readiness asks why a product was filtered out, and the answer is usually something concrete: a price that could not be confirmed, a missing attribute, no independent coverage. We walked through a worked example in I asked ChatGPT for the best running shoes under $150.
How does an agent build its candidate set?
An agent builds its candidate set by gathering options from the sources it has access to, discarding the ones it cannot verify against the question, and keeping the few it can describe with confidence. The details differ by agent and are not public, but from the outside the process looks like three filters.
- Can I find it? The product has to exist somewhere the agent looks: a product feed it ingests, a page it can open, or a source it trusts such as a buying guide or forum thread. Different agents favor different sources. ChatGPT Shopping and Gemini lean on structured feeds, browser agents such as Muse read the live store, and Claude and Perplexity lean on what the open web says.
- Does it fit the question? The shopper gave a budget, a use and perhaps a constraint. The agent needs the price, the relevant attributes and availability as facts it can match. "Wide fit available" has to be stated somewhere as data or plain text, not implied by a photo.
- Can I stand behind it? An agent is making a recommendation in its own voice. It prefers products where the facts agree across sources and where someone independent has said something good. Conflicts between your feed, your page and your reviews are a reason to leave you out.
Each filter maps to work you can do, which is why our services run in the order Audit, Fix, Cite, Track.
Which metrics describe your place in the candidate set?
Three metrics describe your place: candidate-set presence, position, and share of recommendation. They build on each other, and presence comes first because the other two are zero without it.
| Metric | What it asks | How it is counted | What it tells you |
|---|---|---|---|
| Candidate-set presence | Was the brand mentioned at all? | Yes or no, per answer | Whether agents can find and verify you |
| Position | Where in the answer did it appear? | First, second, third and so on | How strongly the agent prefers you once you are in |
| Share of recommendation | In what share of answers was the brand actually recommended? | Recommended answers divided by total answers | Your overall standing in a category, comparable over time |
A passing mention is not a recommendation. "Some shoppers also look at Brand X, though reviews are mixed" counts for presence but not for share. The full scoring rules, including how we handle disagreements between coders, are in the Index methodology.
Is it really the only metric that matters?
It is the only metric that matters first. Revenue still matters more in the end, but in agent-led shopping nothing downstream can happen until you are in the set. Traffic, conversion rate and average order value all assume the shopper reached you. If an agent is doing the shortlisting, presence is the gate in front of all of them.
There are honest limits to this claim, and they are worth stating:
- Presence does not measure sales. It measures what agents say. Treat it as a leading indicator and watch revenue from agent-referred sessions alongside it where you can identify them.
- Agents change without notice. A result is a record of one week. The trend across months tells you more than any single reading.
- Not every category has moved equally. For considered purchases with clear specifications, shoppers already ask agents. For impulse buys, less so. Check your own category before you reorganize a team around it.
Even with those caveats, the ordering holds. A team that knows its presence across a realistic set of questions knows where to work.
How do you measure candidate-set presence?
You measure candidate-set presence by asking a fixed set of realistic shopper questions, repeatedly, across several agents, and recording which brands appear. No agent will report it to you, so it has to be sampled.
- Write the question set. Twenty to forty questions per category, in full sentences, each with a budget, a use and a constraint. Include comparison and "is it worth it" questions.
- Fix the conditions. Default consumer products, signed out where possible, no memory or custom instructions, and a stated location, because answers differ by market.
- Repeat every question. Answers vary between runs. We ask each question five times per agent for the Agent Index.
- Record three things per answer. Mentioned or not, position, and recommended or not. Keep the transcript.
- Track named competitors the same way. Your share only means something next to theirs.
- Re-run on a schedule. Weekly or monthly. Keep most questions fixed so results compare, and rotate a few so you are not tuning to the test.
This is laborious by hand, which is why it is a service. Track runs forty or more prompts per category weekly across five agents and alerts you when a recommendation drops. Pricing is published on the site.
How do you get into the candidate set if you are not in it?
You get into the candidate set by removing whichever of the three filters is excluding you: findability, fit or defensibility. Start with a diagnosis, because the fix for each is different.
- Never mentioned, in any agent: usually a findability problem. Check feeds, crawler access and whether any independent source covers you.
- Mentioned in feed-led agents but not others: usually a citation gap. The open web is not talking about you in the places those agents read.
- Mentioned but rarely recommended: usually a fit or consistency problem. The agent knows you exist but cannot confirm the details the question asked for, or the sources disagree.
For the step-by-step version, read how to get your products recommended by ChatGPT. To find out which bucket you are in today, run the free check or book a readiness call.
Key takeaways
- The candidate set is the three to five products an agent shortlists for a specific question. Outside it, you are invisible.
- It is rebuilt for every question, so presence is a rate across many questions, not a single yes or no.
- Agents filter on three things: can they find you, do you fit the question, and can they stand behind the recommendation.
- Measure presence first, then position, then share of recommendation, always next to named competitors.
- It is a leading indicator, not revenue. Watch the trend, not a single reading.
Questions and answers
What is the candidate set in agentic commerce?
The candidate set is the short list of products or brands an AI agent seriously considers for a specific shopping question before it writes its answer, usually three to five names. Products outside it are never shown to the shopper.
What is share of recommendation?
Share of recommendation is the percentage of answers in which an agent actually recommends a brand when asked for the best product in a category. It is calculated as recommended answers divided by total answers, across repeated runs and several agents. Passing mentions do not count.
How is the candidate set different from a search ranking?
A search ranking is a long ordered list where lower positions still get some visibility. The candidate set is a small group with a hard edge: you are either named or absent. It is also rebuilt for each question, so presence is measured as a rate across many questions.
How many prompts do you need to measure candidate-set presence?
Enough to cover how shoppers really ask. We use forty questions per category for the Agent Index, each asked five times per agent because answers vary. For a first internal check, ten to twenty well-written questions asked three times each will show whether you have a problem.
Written by the team at Recommended by Agents. Published . Spotted something out of date? Tell us and we will fix it.
