The “Inside” series opens the hood on one product at a time, and it has never visited the other shelf — the seven Shopify apps. Time to fix that, starting with the one whose machinery is most worth explaining: Cited, which answers a question merchants increasingly ask and no analytics dashboard can — when a shopper asks an AI assistant what to buy, does my store come up? That storefront is invisible to you. Cited's whole job is to look at it on your behalf, and the honest version of “how” is more interesting than the marketing version, including the part where we tell you what a single run can and cannot prove.
The Measurement Problem
Shoppers ask ChatGPT, Gemini, and Perplexity for buying advice, and those assistants answer with specific product and store recommendations. If your store is in those answers, that's a channel; if it isn't, that's a silence you'd want to know about. The trouble is that you can't check this by feel — asking once from your own account tells you almost nothing, and no report from the assistants exists. What you want is an instrument: something that asks the right questions, the way shoppers ask them, across the assistants that matter, and scores the answers the same way every time. That's the product. Here's the machine.
Read the store — once, quickly
A run starts with a single query to your store's Shopify Admin API: shop details plus your first products — names, types, collections. That's the entire data collection. Cited doesn't crawl your site, doesn't read your customers, and doesn't need to; it needs just enough to know what a shopper would call the things you sell.
Build the questions — deterministically, on purpose
From your product types (or collections, or product titles as a fallback), Cited fills eight fixed buyer-intent templates: best {x} to buy online · best {x} for beginners · affordable {x} under $50 · where to buy {x} · {x} worth the money · top rated {x} 2026 · best {x} brands · {x} gift ideas. These are string templates, not AI-generated questions, and that's a design decision: your trend line is only meaningful if this month's questions are the same as last month's. An AI that creatively rephrased the questions each run would add noise to the exact thing the product exists to hold still.
Ask three assistants at once — in their web-grounded modes
Each run measures four of those questions against three providers concurrently — OpenAI's web-search-enabled model, Google's Gemini with Search grounding, and Perplexity's Sonar — a dozen live, grounded AI calls per run, fired in parallel so the run takes as long as the slowest answer rather than the sum of all of them. If one provider errors — quota, outage — the run completes anyway: its cells drop out of scoring and the report says plainly that it couldn't be reached, rather than quietly pretending.
Why “web-grounded” is non-negotiable
This is the best idea in the product, so it gets its own box. Query a model without live web access and you are measuring its training memory — a snapshot of the internet from months ago, which no shopper experiences. The consumer apps people actually use search the live web before answering. So Cited only ever queries providers in their grounded modes — live retrieval on, citations returned — because the question isn't “did this model's training data include you.” It's “when this assistant goes looking today, does it find you and say so.” Cheaper shortcuts exist and we deliberately don't take them; an ungrounded answer would be a precise measurement of the wrong thing.
Score the answers — without asking an AI to judge an AI
Scoring is deterministic text matching, and we like it that way. Mentioned: your brand name or domain appears in the answer. Cited: your domain appears inside one of the answer's citation links — the stronger signal, because the assistant didn't just say your name, it pointed at you. Prominence: how early in the answer the first mention lands. Roll those up across every question-and-provider cell and you get the run's visibility score — the share of cells where you were mentioned at all. No model grades the results; the same answer always scores the same way. The code labels this scoring “v1” internally, which is accurate — it will miss a creative paraphrase of your brand — and we'd rather ship an honest simple metric than a clever opaque one.
The Paragraph the App Doesn't Say Yet
One run is a sample, not a verdict
Here is the disclaimer that, as of this writing, appears nowhere in the app, the listing, or the product page — so it appears here first. AI answers vary. Ask the same grounded question twice and you can get different answers; Cited asks each measured question once per provider per run, with no repeat-sampling or averaging. And the assistants' consumer apps personalize by geography, session, and history, so the API's answer approximates — but does not perfectly equal — what any particular shopper sees. What this means practically: a single run is a sample. Don't celebrate one good run or panic over one bad one; the signal is in the trend across runs, which is exactly why the product keeps history. We build measurement tools, and a measurement tool that overstated its own precision would be defeating the purpose. This paragraph is the app's missing label, and it's on the record now.
What It Refuses to Do
The Other Shelf piece gave every app one refusal, and Cited's stands unchanged: it measures; it does not persuade. Nothing in this machine injects your store into a model, buys placement, or “optimizes” an assistant's opinion of you — and anyone selling that service is selling something that doesn't exist to sell. What the visibility picture is for is your own strategy: product data, positioning, content. Two smaller honesty notes while the hood is up. Cited runs on demand — you click, it measures; there is no background scheduler quietly burning your run allowance, and no run happens without you. And the trend line is only as long as your clicking makes it, which is a genuine limitation of the on-demand design, not a feature we're dressing up.
What It Costs, and Why the Cap Exists
Free runs the real instrument — 3 full measurement runs every 30 days, no card. Pro is $29/month or $290/year with a 7-day trial for first-time subscribers, and paid stores get a cost-guard ceiling of 20 runs per hour — which sounds like a strange limit to advertise until you remember each run fires a dozen live web-grounded AI calls on our bill. The ceiling exists so the pricing can stay flat instead of metered; normal use never touches it.
The Honest Summary
Cited is a small instrument built around three convictions: ask stable questions (templates, not improvisation), ask them where the truth lives (web-grounded modes, never training memory), and score the answers so simply that nobody — including us — can argue with what a number means. Around that sits the honesty the instrument owes you: providers fail visibly, the scoring admits it's v1, and one run is a sample rather than a verdict. That's the whole machine. If the question “do the AI assistants know my store exists?” has been nagging at you, it's answerable now — measured, not guessed. For how a sibling product uses AI in almost the opposite way — generation rather than measurement — the PostPilot deep-dive is the companion read, and the AI-tooling essay covers where we think this all heads.
Ask the Question Properly
Three assistants, twelve grounded calls, one honest score — free tier included, no card required.
Explore Cited