On October 26, 2026 we asked Shopify’s agent catalog for a “humidifier for dry winter air large room quiet top fill” and read the ten listings it returned, field by field. Four product descriptions gave a decibel figure. Two more listings had one only in metadata that Shopify generates. No title had one, and not one figure said which setting, at what distance or by what method it was measured.

So the listing that says “21dB Quiet” gives an agent something it can compare, and its neighbour that “runs quietly enough to leave on overnight” doesn’t. That’s a difference in evidence. It doesn’t show that either product is quieter, or that either was rejected or recommended more. This is one catalog snapshot, not a live ChatGPT result.

The two listings, side by side

These two sat in the same top 10. The quotes are exact; the store names are withheld.

Fact Position 2 Position 4
Noise, in the description “21dB Quiet” “runs quietly enough to leave on overnight”
Room size, in the description “Large Room 700ft²” “rated for single rooms up to 400 square feet”
Runtime, in the description “100H”, no setting “up to 36 hours on low”
Shopify-generated metadata Spec list present, no noise line None in the record

On noise, position 2 is more specific. On runtime it’s the other way round: position 4 names the setting its hours apply to, which only one other description on the shelf does. Specificity belongs to each fact, not to the whole listing, so you have to check fact by fact.

They aren’t the same product either. Position 2 lists an 11-litre tank at USD 119.99, and position 4 a 1.25-gallon tank at USD 69.99. A shopper choosing between them would be weighing more than noise.

Every noise statement on the shelf, by field

This is the full audit for the ten results. “Description” is the product’s description text in the catalog response. “Generated” means Shopify’s spec list, feature list and selling-point line (metadata.tech_specs, metadata.top_features and metadata.unique_selling_points).

Position Description says Generated metadata says What that leaves you with
1 Nothing about noise “Quiet operation ensures minimal disruption”; “quiet” in the selling point Quietness only in generated text, in words
2 “21dB Quiet” No noise line in the spec list The number is in the description only
3 “At just 26 dB” “Noise Level: 26 dB”; “Ultra-quiet operation at 26 dB” Consistent, no condition
4 “runs quietly enough to leave on overnight” No generated metadata in this record Words only, no number anywhere
5 “brings noise down to 28dB” “Noise Level: 28 dB”; “Hypersilent operation at 28 dB” Number consistent; “hypersilent” is generated
6 Nothing about noise “Noise Level: As low as 26 dB”; “whisper-quiet” Number only in generated text, with “as low as”
7 “Ultra-quiet 26 dB operation” “Noise Level: 26 dB”; “Ultra-quiet 26 dB operation” Consistent, no condition
8 “ultra-quiet making it ideal to have on while you sleep” No noise line Words only
9 “low noise level” “Noise Level: 26 dB”; “Ultra-quiet operation at 26 dB” Words in the description, number only in generated text
10 Nothing about noise Nothing about noise No noise statement in any field we read

Every count below names the field it counts:

  • Title: a decibel figure in 0 of 10.
  • Description: a decibel figure in 4 of 10 (positions 2, 3, 5 and 7: 21, 26, 28 and 26 dB).
  • Description or generated metadata: a decibel figure in 6 of 10 (adding positions 6 and 9, both 26 dB, generated only).
  • Description, noise in words only: 3 of 10 (positions 4, 8 and 9).
  • Description, no noise statement at all: 3 of 10 (positions 1, 6 and 10). Generated text adds one for 1 and 6; position 10 says nothing about noise anywhere.
  • A measurement condition with any decibel figure: 0 of 4 in descriptions. The only qualifier anywhere is “as low as”, in generated text for position 6.

The shopper also asked for top fill. “Top fill” appears in 7 of the 10 descriptions (positions 1 to 7). Position 9 says it only in its title, so 8 of 10 say it in the title or description.

What the generated fields added, and what they dropped

Shopify says the fields it generates or enriches are marked Inferred, might not always be present, might vary in accuracy, and should be treated as discovery signals rather than merchant-authored text. On this shelf you can see why that label matters:

Position Description Generated metadata Change
2 “21dB” No noise line The most specific noise figure on the shelf isn’t in the generated spec list, which does list two items as “Not specified”
6 No noise statement “As low as 26 dB” A number appears that the description never states
9 “low noise level”; “up to 25 hours” “26 dB”; “Up to 25 hours (low mode)” A number and a runtime condition appear
5 “up to 32 hours” “Up to 36 hours” The runtime changes
1 No noise statement “Quiet operation” A quietness claim appears, in words

We can’t tell where the added values came from. Shopify may have used product fields, manuals or other merchant data that the response didn’t include, and the 26 dB for positions 6 and 9 may well be correct. The point isn’t that generated metadata is wrong. It’s that a fact you care about should be in text you wrote, in a field you control, so that anything reading your record finds the same value with its condition.

Shopify also marks the description field itself as generated or enriched. So even the description column above isn’t guaranteed to be the merchant’s own wording. We quote what the catalog returned.

What a decibel figure doesn’t tell you

A bare “26 dB” answers “how quiet?” with a number, which beats an adjective. It still leaves three questions open:

  1. Which setting? A humidifier on its lowest mist level is quieter than on its highest. Only position 6’s generated text hints at this, with “as low as”.
  2. How far away? Sound level falls with distance. None of the ten records names a measuring distance.
  3. Which scale and method? “dB” alone doesn’t say whether the figure is A-weighted or how it was tested. None of the ten records says.

So when an agent compares 21 dB with 26 dB, it’s comparing two numbers whose conditions it can’t see. They may have been measured the same way. Nothing in the record says so.

The runtime figures have the same gap. Eight of the ten descriptions give a runtime, and only two say which setting it applies to. One description lists a 6.2 L tank, 320 ml/h of mist and up to 62 hours. By our arithmetic, 6.2 L at 0.32 L/h lasts about 19 hours, so the 62-hour figure implies a lower average output that the description doesn’t name.

What this does and doesn’t show

It shows what each listing supplied for a question a shopper actually asked. It doesn’t show:

  • Which humidifier is quieter. We measured nothing. Every figure is a seller or Shopify statement.
  • That the adjective listing was rejected. We can’t see what any assistant did with this shelf. Shopify’s catalog is one source assistants can use, and a catalog response isn’t a ChatGPT, Copilot or Gemini answer.
  • That adding a number lifts recommendations. Nothing here measures lift.

The closest evidence we have is a simulation. We placed an invented humidifier among seven real listings and asked six models to pick two of eight. With vague copy, it was picked in 3 of 180 Claude sessions and between 0 and 6 of roughly 60 sessions on each GPT and Gemini model, far below the 25% that random picking would give. With its facts stated, it was picked far more often. On GPT and Gemini, though, it won only when the shopper’s brief asked for something its facts answered and the other listings didn’t. When competitors already stated the same or quieter decibel figures for less money, a number alone didn’t win. The simulation write-up has the detail. It’s a supplied-snippet test with an invented product, not live shopping.

How we checked

We sent the query “humidifier for dry winter air large room quiet top fill” to Shopify’s public Global Catalog through Shopify’s own command-line tool (catalog search, top 10, no buyer, country or currency context, no cart or checkout) on October 26, 2026. The response used protocol version 2026-08-25. We saved the response exactly as received.

For each of the ten products we read:

  • title;
  • description (we stripped the HTML for reading; each product’s only returned variant carried the same description);
  • metadata.tech_specs, metadata.top_features and metadata.unique_selling_points.

We counted a decibel figure wherever a number was followed by “dB”, and read each noise word in context. We checked every quote and count again against the saved response on October 27, 2026. That was a re-read, not a fresh catalog query. Results can change by day, country and buyer.

We name listings by position, not by store. Six of the ten came from one seller, so this shelf partly reflects one company’s writing habits. It isn’t a sample of humidifiers. This page replaces our shorter October 26 note on the same contrast, which quoted the same two listings.

Audit your own record

You can run the same audit on one of your products in about ten minutes. A developer can pull your catalog record with Shopify’s UCP command-line tool (Shopify’s Global Catalog documentation shows its catalog search and catalog lookup commands). Or start from your product page and admin fields.

Field What to look for The warning sign on this shelf
Title The one number a shopper filters on, if it fits No title gave a decibel figure
Description Value, unit and condition together, for example “[x] dB(A) on low, measured at [distance]” Four figures, no conditions
Generated spec list Does it match your description? Is anything missing or changed? 21 dB missing; 32 h became 36 h
Generated features and selling points Claims you didn’t make “Quiet operation” on a listing whose description never mentions noise

The record you’re building toward is the one in our product-record guide: one row per fact, with its value, unit, condition and source. Shopping for a humidifier rather than selling one? The humidifier field guide covers what to ask for. For the whole snapshot this shelf came from, see our catalog research.

Your next step

Pick your best-selling product and the one question customers email about most. For a humidifier, that’s how loud it is. Find the number in your test report or manual, with its setting and distance. Then check that it appears, with that condition, in your own description. If there’s no test report, write “noise level: not tested” in your records rather than inventing a figure, and don’t let a generated “ultra-quiet” stand in for one.