Which of Your Products Are Invisible to AI? Measuring Citation Coverage, Not Just Pages

Short answer: most stores do not have a visibility problem, they have a coverage problem. You have 400 products and the assistants can speak confidently about 30 of them. The other 370 are not ranking poorly — they are absent from the answer space entirely, and no single-page fix will reveal that, because you cannot see a gap by looking at pages one at a time.

This is the difference between diagnosing pages and diagnosing a catalogue. Page-level audits answer "is this page citable?" Catalogue-level coverage answers "of everything I sell, what share of it can an assistant actually talk about, and what do the invisible ones have in common?" The second question is the one that changes your roadmap, and almost nobody is asking it.

## Why per-page audits hide the real gap

A per-page audit has a selection bias built in. You tend to run it on pages you already suspect, or on your best sellers, or on whatever traffic data told you matters. That means you audit the products that are already somewhat visible, fix them, and conclude your work is done while three quarters of your catalogue stays dark.

There is a second reason. Citation behaviour is uneven by category. A product type with dense, checkable attributes — connectors, consumables, replacement parts, anything with part numbers — gets cited easily because the facts are unambiguous. A product type whose value is aesthetic, seasonal, or subjective gets cited rarely no matter how well the page is written, because no sentence on it answers a deciding question. If you sample pages, you will sample one kind and generalise wrongly.

## The five signals that predict whether a product can be cited

Before you fix anything, sort the catalogue by these. Each is checkable in bulk rather than one page at a time.

1. Deciding-fact density. How many buyer constraints does the page answer as plain text: fit, compatibility, capacity, material, power, care, warranty term? Products with three or more checkable facts get pulled into answers. Products with zero are decorative. 2. Uniqueness of the facts. Is any of this text exclusive to you? A page that only repeats what every reseller of the same SKU says gives the assistant no reason to prefer your URL over the twenty others. Original measurement, testing, or fabrication detail is what makes your page the source. 3. Question coverage. Which of the questions your support inbox receives in this category are answered on the page? Products where the top questions are answered in text are citable. Products where they are only answered by a support agent are not. 4. Machine-readable completeness. Are the constraints also in structured fields with units, or only in prose and images? Prose is a hint; a field is a fact. The two disagreeing is worse than either alone. 5. Cross-surface consistency. Does the same value appear identically on the product page, the feed, and any marketplace listing? Where they diverge, the assistant has to choose, and choice is the beginning of doubt.

## A practical procedure: build the coverage grid in one pass

You do not need a crawler. You need a spreadsheet with one row per product.

1. Export the catalogue with SKU, category, price, and URL. 2. Add one column per signal above. Score each product 0-2 on each of the five. Bulk-assign from your product data where you can, and spot-check a sample to calibrate your scoring. 3. Rank by total score and look at the bottom quartile. That is your invisible shelf. 4. Group the bottom quartile by cause, not by product. You will usually find three or four recurring causes — a category with no attribute fields, a supplier feed that arrives with empty fields, a page template that renders specifications as an image. Causes fix in batches; products fix one at a time. 5. Re-score monthly. Coverage is a moving number. A template fix moves hundreds of rows at once; that is the payoff you are looking for.

## What the bottom quartile usually turns out to be

In most catalogues the invisible shelf is not random. It clusters into recognisable groups: accessory and spare-part SKUs that were uploaded as a title and a price with no attributes; seasonal items that were never revisited after the season ended; long-tail variants that inherited a parent description which does not actually describe them; and category pages for product types where nobody on the team knows which facts buyers need, because the questions were never recorded.

That last one is the most valuable finding, because it is a process failure rather than a page failure. If nobody has ever written down the deciding questions for a category, no amount of copywriting will guess them correctly.

## The part that is genuinely incremental

Fixing the naming or the schema on an invisible product does not guarantee it gets cited. Nothing does — assistants weigh sources they cannot fully control, and you are competing against pages that may hold stronger evidence than yours. What coverage measurement gives you is that it stops you from working on the wrong pages, and it converts "improve our GEO" from an endless editorial project into a ranked list with a denominator: 370 products to bring into the citable set, grouped into four causes.

You can see this shape of result for your own catalogue without building the spreadsheet first. A free scan at /audit reports, per URL, which facts exist as extractable text, which buyer questions go unanswered, and where page values conflict with the feed. It is a diagnostic, not a guarantee — but it is a fast way to find out whether your gap is a handful of pages or a whole shelf. If you want the underlying reasoning for why extractable facts behave differently from visual ones, the /faq page walks through it.

## FAQ