ai

How to Test Whether Your Range Is Readable by AI Before You Pay Anyone to Fix It

A DIY audit before you buy anything: a product sample, shopper questions, several assistants, a scoring sheet. What to record, what it can't prove, when a tool helps.

F
Furniture Connect TeamNovember 18, 2026
How to Test Whether Your Range Is Readable by AI Before You Pay Anyone to Fix It

The one-off test in ask an AI assistant for a sofa gives you an initial observation. Before you spend money on a tool, you want something more repeatable: where you stand, on which products and on which questions. Here's a method that needs nothing but a spreadsheet and the assistants you want to test. The time depends on the sample, so start small enough to finish and repeat.

1. Pick the sample

Choose products deliberately rather than at random. Include best-sellers, a recent launch, products with thin pages and products that compete directly with brands you have seen recommended. Note each product's name, URL and the variant you will treat as the default.

Start with a sample you can finish, then widen it once the method works.

2. Write the questions

Two kinds.

Shopper questions, phrased as a shopper would, that don't name your brand: "a sectional under 100 inches for a small apartment with a washable fabric", "a leather sofa under $2,500 that ships in three weeks". Eight to ten of these, covering fit, materials, durability, delivery and price. The list in the questions shoppers ask AI assistants about sofas is a start; your customer service team's top questions are better.

Named questions, about your products directly: "tell me about the Harlow sofa from [brand]", "what's the seat depth of the Harlow", "does the Harlow fit through a 30-inch door". Three or four per product.

The two kinds measure different things. Named questions tell you whether an assistant can read your product when it's pointed at it. Shopper questions tell you whether it reaches for your product unprompted, which is the one that matters commercially and is much harder.

3. Choose the assistants

Use more than one. They differ in how they search, what they read and whether they use product feeds, and a result from one isn't a result from all. Three or four of the widely used ones is plenty. If you sell in more than one country, run from each, because answers vary by region.

4. Run it and record it

For each question on each assistant, record:

  • Which products and brands were recommended (shopper questions).
  • Whether yours was among them.
  • For named questions: which facts the assistant stated correctly, which it got wrong, and which it said it couldn't find.
  • Which sites it cited or linked, if it shows them.

Keep the raw answers. The "couldn't find" column and the "wrong" column are the most useful things you'll produce; they're a list of specific gaps and specific errors, per product.

Do it on one day, and note the date. Assistants change; a baseline is only useful if you know when it was taken.

5. Read the results

Three questions to ask of the sheet.

Which facts are consistently missing? If seat depth is "couldn't find" across fifteen products, check the record and the published page before concluding. It may be an attribute gap, a publication gap, or the assistant failing to retrieve the page or match the variant. If the attribute is filled and on the page as text, the problem is downstream of your catalog.

Which facts are consistently wrong? If width is wrong and always matches your default variant, one possibility is that your variant data isn't reaching the page as text. If materials are wrong, check whether your page says something vague that the assistant filled in.

Who's recommended instead, and what do their pages have? Open two of their pages and do the side-by-side from what the AI found on your competitor's page that it didn't find on yours. Differences in what each page states as text are often visible in the specifications table; they don't on their own explain the choice.

6. Decide what to fix first

Sort the gaps by how many products they affect and how cheap they are to close. A fact that exists in a PDF but not as text is cheap and wide. An attribute nobody has ever recorded is expensive but might affect every product. Fix the catalog-level items first: create the attribute, add it to the relevant family, fill it across the range and make the template show it. Then re-run the same sheet and compare.

What the test can't tell you

It can't tell you why an assistant chose what it chose; you infer that from what it could and couldn't read. It can't tell you how many shoppers are asking; it tells you what happens when they do. It's a sample, taken on a day, and assistants vary from run to run, so a single result isn't precise. Treat a change between two runs as a signal when it's large and consistent, not when one answer moved.

And it does not scale by hand. Add more products, questions, assistants and regions and the number of answers to record grows quickly. Repeating the exercise after significant catalog changes turns a useful pilot into recurring manual work.

When a tool earns its place

Once you've run this once by hand, you know exactly what you'd want automated: the same shopper and named questions, asked across the assistants in your regions whenever you re-run it, with the results scored per product and the "couldn't find" column turned into a list of empty attributes on the record to fill. If a tool does that, and connects the gap to the fix (the attribute, the image, the field), it's doing the work you just did, on request. If it reports a visibility score without telling you which fact on which product is missing, it's telling you there's a problem, which you already know.

Furniture Connect has an AI visibility check built along those lines. The working method and tradeoffs are in how to run an AI visibility check on your range. Whatever you use, run the manual version first. It makes you a much better judge of the automated one.

For the whole argument, see being found by AI.

This week

Build the sheet. A practical pilot is five products, five shopper questions and two assistants. Widen it once you know what you will do with the results. The "couldn't find" column is your plan, once you've checked each gap against the record and the page.

Ready to show and sell?

Give every team the product data, imagery and assets they need to launch faster, quote better and turn more of the catalog into revenue.

No slides. We stage one of your products live, at a time you pick.

Talk to salesOr see pricing