Pulse

Know which photos prove what buyers need

Visual Evidence checks product photos against buying requirements, revealing what they prove, what contradicts your data, and what they can't show—so you know which photo to add before updating product fields.

A laptop's photo gallery, one side photo under a magnifying glass, joined to a card of buying requirements: ports shown in the photo, colour contradicting the data, the port side needing a photo, and battery life, which no photo can show.
A complete gallery of laptop photos marked gallery complete, beside a shopper with a question: the port side is still a question mark, and battery life is marked as something no photo can show.

A complete gallery can still leave the question open

A laptop shot from every angle can still never show the port side. A sofa photographed perfectly can still leave a buyer unsure it fits through the door. A gallery checklist counts photos; it does not ask whether any photo answers the question the buyer is stuck on.

And some facts no photo can carry, battery life or a weight. Asking an AI to read them off a picture is how a storefront ends up confidently wrong.

The right answer is sometimes: no photo can show this. Record it instead.

How does Visual Evidence work?

The same export as the Decision Audit, with your image links. Nothing goes on your site.

  1. A spreadsheet of products, categories and image links sliding into an upload tray, with a badge saying no customer data.

    01You

    Send one category with its images

    Your catalogue export for one category, with the image links you already publish. No customer data.

  2. A laptop's side photo under a scanning frame, passed to the requirement engine, which marks ports as shown in a photo, soldered memory as no photo, and battery life as never.

    02Us

    We check each photo against the buying requirements

    For each requirement in the category, which facts it needs and whether a photo can show each one: ports yes, soldered memory no, battery life never. A vision model reads what each photo shows; our requirement engine, not the model, decides what it means.

  3. A ranked list of repairs: first add a photo of the port side, then record the ports, then reconcile the colour, linked to the photos and the data they change.

    03Both

    You get the photo and data repairs, ranked

    The facts a photo shows but your data lacks, the photos that contradict your data, and the photo to add for a fact that matters, alongside the Decision Audit's field repairs.

What you get

The ports in a laptop's side photo lifted into the data, where the ports row is now checked while soldered memory and battery life stay empty.

Facts your photos show but your data lacks

A port, a colour, a dimension diagram already in your gallery and missing from the field a filter or an AI assistant reads. Lift it into the data.

A data card saying colour navy and two HDMI ports, contradicted by a photo of a green laptop with one HDMI port, beside a returned parcel.

Photos that contradict your data

The spec says two HDMI ports and the photo shows one; the colour field says navy and the photo is green. A buyer who trusts the photo returns the product.

A laptop gallery with an empty slot for a photo of the port side, linked to a desk where the laptop drives two monitors.

The photo to add, where it decides

Add a photo of the port side for these laptops, because buyers choosing for two monitors depend on it. A photographer's job, not data entry.

Battery life, weight and upgradeable memory, each marked as something no photo can show, all leading to one place: record it.

What no photo can show

Battery life, weight, memory that cannot be upgraded: marked as facts to record, never read off a picture.

An AI chip saying cannot tell, beside a benchmark scorecard where cannot tell scores higher than a lucky guess.

A model that says it cannot tell

Every model is scored on our benchmark before it reads your photos, and a correct cannot tell scores higher than a lucky guess.

What Visual Evidence will and will not do

What this does

  • Say which photos show, contradict or cannot show each fact your buyers need
  • Recommend the photo to add, ranked with your data repairs
  • Say cannot tell when no photo can show a fact, rather than guess
  • Show you the benchmark a model passed before it reads your photos

What it does not do

  • Promise more sales or fewer returns. It has not run on a merchant's catalogue yet, so there is no measured result to quote
  • Invent a fact from a photo. A value no photo can show is reported as missing, not estimated
  • Generate or edit images. It tells you which photo matters; your team or studio takes it
  • Replace the Decision Audit. It adds photos as evidence to the same measurement

The ones that decide it

Is it available now?

Early access. It is offered first inside a Decision Audit pilot, on one category. It is built and benchmarked on public product data, and has not yet run on a merchant's own catalogue.

How do you know the model does not make things up?

Each model is tested first on a benchmark of product questions, including ones no photo can answer. In our first run on public product data, two models answered cannot tell 98% of the time when no photo could show the fact. That is public data and two models, not your catalogue and not a promise.

How is this different from checking a gallery against image requirements?

A gallery check asks whether you have the required number and type of photos. Visual Evidence asks whether any photo proves what a buyer needs to establish, and whether it agrees with your data.

Do you create the missing photos?

No. We tell you which photo to add and why. Your team or your studio takes it, and the next audit confirms it shows what it should.

What data do you need?

One category export with the image links you already publish. No customer data and no access to your systems.

Start with one category and its photos

Ask for Visual Evidence with a Decision Audit pilot.

Request early access

Early access, priced in the pilot.