When an AI assistant compares your product with three others, every fact it has to dig for is a step it might not take.
Put the fact a shopper is asking about in your title or the first lines of your description, and comparing you costs the assistant nothing. Leave it in a size chart image, on one variant, at the bottom of a long description or on another page, and the assistant may compare you without it.
That’s fine when the missing fact doesn’t matter. It’s a lost sale when it’s the one fact the shopper asked about.
Every lookup is a step an agent can skip
A shopping agent works on a budget, like a person does. Every page it opens and every product it looks up in detail costs time and money for whoever runs it. Shopify’s own tools for agents work this way: an agent searches the catalog first, then makes a second call for each product it wants to look at more closely.
Researchers at McGill measured what happens when looking things up has a cost. In their Tool-Lab study, eight models from Google, OpenAI and Anthropic chose between two coffees: $4.99 for 10 oz and $5.00 for 11 oz.
The second is the better deal per ounce. But each fact (dollars, cents, weight) stayed hidden until the model paid to look it up.
Here’s what happened:
| The shopper said | Lookups free | Lookups cost $10 each |
|---|---|---|
| “Find the coffee with the lowest price per ounce” | 94% picked the better deal | 83% |
| “Find the best deal on instant coffee” | 89% | 29% |
With a clear request, the models paid to check what they needed. With a vague one, they checked the dollars and skipped the cents and the weight, then picked the cheaper-looking bag. The authors found most of the wrong picks came from missing information, not bad math.
Now think about how your customers actually talk to ChatGPT. “Best humidifier for a big room” is a vague request. “Find the humidifier with the lowest noise in decibels for rooms over 400 sq ft” isn’t something anyone types.
Facts that are read first get compared
The other half of the problem is order. AI assistants don’t always read your whole description. OpenAI says ChatGPT may write its own simplified titles and descriptions from the information it receives.
A second research team saw the effect in a mock marketplace. In the ACES experiments, a floor lamp’s listing began “Floor Lamps for Living Room” and only said “Office” at the very end. The researchers think that on their platform, which cut descriptions short, buying agents never saw the word when a shopper asked for an “office lamp”.
When the rewrite moved “Office” to the front, the lamp gained share with all six buying models tested. It was the only product category where the rewrite helped every model.
Our guide to clipped product descriptions covers how to write for the cut. The short version: lead with the facts.
Where facts hide on a Shopify listing
Here are the places a fact can live, from easiest to hardest for an assistant to reach, with examples:
| Where the fact lives | How hard it is to reach | A real example |
|---|---|---|
| The title or first lines of the description | Read first, every time | A snowsuit’s title: “protects to -30 °C” |
| Further down a long description | May be cut or rewritten | A floor lamp that said “Office” only at the end |
| Only in Shopify’s generated lists | There, but marked “Inferred”: possibly missing or wrong | A humidifier whose description says “low noise level” while only Shopify’s list says “26 dB” |
| Only on a variant | Missed when the product-level record is read | A lip gloss gift box whose product description was blank; the full text sat on the variant |
| In an image, a PDF or a size chart picture | Needs an extra fetch, and may not be readable as text | A size chart saved as a picture |
| Nowhere | Can’t be compared at all | That same humidifier never states a room size |
Shopify labels its generated fields “Inferred” and says to treat them as hints, not facts you wrote. So a number that only appears there is a step removed from you, even if an assistant can see it.
Count the steps on your best seller
This takes 15 minutes. Take your best seller and write down the five questions customers ask about it most. For each one, find the fact that answers it and note where it lives, using this scale:
- 0 steps: in the title or the first two lines of the description, as a number with a unit.
- 1 step: further down the description.
- 2 steps: only on a variant, only in Shopify’s generated lists, or only on another page.
- 3 steps: only in an image, a PDF or a size chart picture.
- Not there: nowhere a machine can read it.
Here’s the humidifier from our one-product audit, scored this way:
| Shopper question | The fact | Where it lives now | Steps |
|---|---|---|---|
| Is it top fill? | Top fill | Title: “2.5L Top Fill” | 0 |
| Is it quiet? | 26 dB | Description says “low noise level”; the number is only in Shopify’s list | 2, and vague where it’s read first |
| Will it handle a big room? | Room size | Nowhere | Not there |
Your goal: get the fact behind your top two questions to 0 steps. The rest can wait.
Test it on your own product
Want to see the effect for yourself? There’s a quick version you can run in an afternoon.
- Pick one of your products and three competitors that show up for the same search. Copy each one’s title and description into a document.
- Make two versions of yours. Version A: the deciding fact in the first line, as a number with a unit. Version B: the same fact, but only in the last paragraph, or only in a “specs” note under the description.
- In a fresh chat with an AI assistant, paste the four listings and ask a vague question (“Which of these is best for a big, quiet bedroom?”). Then try a specific one (“Which is the quietest, in decibels?”). Do each ten times, for each version.
- Count two things: how often your product gets picked, and how often the answer states your fact correctly.
This tests how a model reads, not how ChatGPT ranks live results. For the live side, ask an assistant with web search to compare your product and one competitor on that fact, three times, and check whether it gets your number right and which pages it cites. To turn a fix into a proper before-and-after, use our testing guide.
What to fix first
- Put the deciding fact up front. First line of the description, number and unit, how it was measured: “26 dB on the lowest setting”.
- Fill in the product-level description. Don’t leave it blank and rely on variant text.
- Turn image-only facts into text. If your size chart is a picture, add the key measurements as text too. Our size chart guide shows how.
- Don’t lean on Shopify’s generated lists. If a number there is right, write it in your own copy, backed by your own documents.
- State your limits. “For rooms up to 300 sq ft” is a fact an agent can compare. Our guide to limits worth publishing has more.
How we checked
On November 4, 2026 we read the Tool-Lab and ACES papers in full and re-read Shopify’s Global Catalog documentation and OpenAI’s shopping help page. The listing examples come from Shopify Catalog results we saved on October 26 and 28, 2026, read field by field. The steps scale is our own rule of thumb.
Your next step
Score your best seller’s top five questions with the steps scale above. Pick the one with the most steps that customers ask about most, and move that fact into the first line of the description this week.