Traffic acquisition Seller Tales

Can You Measure AI Recommendations? A Field Guide to AI Share of Voice

Full disclosure: Seller Tales is published by JoyCraft — when we recommend our own tools below, we say so.


Shoppers are asking ChatGPT, Gemini, and Perplexity which store to buy from. Some stores get recommended. Most don’t know either way — because AI answers leave no report. There is no “AI traffic” tab in your analytics, no dashboard counting “mentions this week,” and no way to log in and see which buyer prompt you just won or lost.

That doesn’t mean the channel is unmeasurable. It means the measurement is yours to run. This guide shows a manual method to measure whether AI assistants recommend you, what you can actually fix, and how to verify a fix worked — the same loop a tool like AisleLens automates, minus the tool.

Full disclosure: Seller Tales is published by JoyCraft — when we recommend our own tools below, we say so.

Why AI recommendations are invisible — and why that matters

Three things make AI recommendations hard to track, and they’re structural, not accidental:

  • Answers vary. The same question asked twice can produce two different answers. AI is non-deterministic; a single “you got mentioned” result proves nothing.
  • There’s no report. Google gives you Search Console. Amazon gives you a referral dashboard. ChatGPT, Gemini, and Perplexity give you a chat box. Attribution stops at “someone opened a link.”
  • Sessions don’t get attributed. Even when a shopper finds you through an AI answer, the visit looks like direct traffic. You can’t tell it apart from a typed URL.

The consequence is that most stores operate on vibes: “I think we show up in AI answers” or “I’ve never seen us there.” Both statements are guesses. The fix is a small, repeatable measurement you can run in an afternoon — and re-run after any change.

AI share of voice: the number you can actually measure

The measurable version of “does AI recommend me?” is share of voice: the share of AI answers to relevant buyer questions in which your store is mentioned, versus your named competitors.

AI share of voice = mentions ÷ total answers tested

Run 10 questions, your store is mentioned in 3 answers → you have roughly 30% share of voice on that prompt set. It’s not a precise market metric; it’s a comparison metric — your number today vs your number after a fix, and your number vs a competitor’s number under the same prompts.

That’s the key discipline: always compare like-for-like. Same prompt set, same engines, same day, same number of runs. Only then does the number mean anything.

A manual measurement you can run tonight

1. Build a prompt set

Use the questions your real buyers type. Two kinds:

  • Category queries (“best inventory management app for Shopify,” “where to buy ceramic dinnerware that ships fast”) — these are the shopping-discovery moments.
  • Comparative queries (“X vs Y,” “alternatives to Z”) — where a competitor gets mentioned and you don’t.

Aim for 8–12 prompts. Write them down before you start, and reuse the exact same wording on every run — changing the wording changes the answer.

2. Run each prompt on each engine

ChatGPT, Gemini, and Perplexity each give different answers, so track them separately. Run each prompt a few times per engine (3–5 runs is a reasonable minimum sample for a first pass — answers vary, and one run lies). On each run, record:

  • Your store mentioned? (yes / no)
  • Any named competitor mentioned? (and which)
  • Which engine, which prompt, run number

Keep it boring and mechanical. A spreadsheet with columns — date, engine, prompt, run, you-mentioned, competitor-mentioned — is the whole tooling you need.

3. Score it

For each engine: your share of voice = your mentions ÷ total runs. Then break it down by prompt: which prompts do you win, which do you consistently lose? Those losing prompts are your fix list.

One honest caveat: manual runs are your own sampling, not a statistically airtight panel. What manual measurement is great at is direction and change — spotting that a competitor appears in 8 of 10 answers while you appear in 2 is actionable even if the exact ratio wobbles.

What actually makes AI recommend a store

When you find prompts you’re losing, the fix list is shorter than most sellers expect. From what’s known publicly about how these assistants answer shopping questions, the levers are the same content signals search engines have rewarded for years — with more weight on clarity:

  • Product titles that match how buyers ask. If shoppers ask “reusable coffee cups 12 oz,” a title that says “Reusable Coffee Cup” reads better to an AI than a title that says “EcoTumblr Pro 2.0 (2026 Edition).” Plain, descriptive, keyword-first titles get picked up in answers.
  • Product descriptions that answer the question. AI answers are built from sources it can read and trust. A description that states size, material, and use case in clear sentences is more useful to an answer generator than marketing copy full of claims.
  • Structured data (JSON-LD). Marking up products, prices, availability, and reviews tells assistants what a page is. It’s the difference between “a page about something” and “a page about this product at this price, in stock.”
  • FAQ and comparison content. A store page that explicitly answers “how is X different from Y” gives an AI a ready-made citation for comparative queries — which are exactly the prompts where you’re probably losing.

None of this requires an AI tool. It requires reading your losing prompts and answering them on your own pages, clearly.

Fix, then verify — the loop that makes it real

The measurement only pays off if you re-run it after a change. The loop:

  1. Baseline. Measure before you touch anything. Write down the numbers.
  2. One change at a time. Fix one thing — say, structured data on your top 10 products, or rewriting three losing product titles. Changing five things at once means you can’t tell which one moved the number.
  3. Re-test the same prompts. Same engines, same wording, same day of week if you can. Compare like-for-like.
  4. Judge by direction, not by one run. One run that mentions you after a fix proves nothing; 7 of 10 runs mentioning you after 3 of 10 before is a signal.

Two cycles of this is usually enough to see which of your pages AI assistants read and which they don’t — which is exactly the information a visibility tool formalizes.

When a tool makes sense

Manual measurement works and it’s free. It also has a ceiling: sample sizes are small, you only re-test when you remember, and you never see prompts you didn’t think of. That’s where a tool earns its keep — running a larger, structured sample across engines, flagging your weakest engine, and tying losses to the exact prompts.

We build one of those tools, so full disclosure: AisleLens (free plan) does this for Shopify stores — it measures your AI share of voice across ChatGPT, Gemini, and Perplexity, shows which buyer prompts you’re losing, and verifies fixes with before/after stats. Start with the manual method above to understand the loop; if the loop becomes part of how you run the store, that’s when the tool is worth a look.

Bottom line

AI recommendations are becoming a real sales channel, and you don’t get a dashboard for it. The workaround is a measurement you control: build a prompt set from real buyer questions, run it across engines a handful of times, score your share of voice, fix the clearest content gap, and re-test the same prompts. It’s an afternoon of work that turns “I think AI recommends us” into a number you can act on — and the discipline of comparing like-for-like keeps that number honest.

More tools from JoyCraft

Independent tools, one small team — by JoyCraft.

Seller Tales is a publication by JoyCraft. When we link to our own tools, we say so, and some of those links carry tracking parameters.