Home › Research › How AI decides who to recommend
SEMPITE Research · Measurement Study · September 2026
Can AI See Your Images? We Checked 2,000 Websites. The Most Visual Ones Scored Worst.
AI answer engines and multimodal models are starting to read images — but only through the machine-readable signals around them. We crawled the images on 2,181 sites across ecommerce, local-service, and personal/portfolio, and read each the way a non-rendering crawler does. The sites whose whole value is their images describe them to a machine the worst.
What we did
We built three cohorts — ecommerce (Shopify storefronts), local-service businesses (dentists, lawyers, home services, salons, gyms, CPAs, restaurants), and personal / portfolio sites (photographers, designers, artists, musicians, authors, coaches, developers) — 2,181 sites in all. We fetched each the way an AI crawler does: static HTML, no JavaScript executed. For every reachable site we measured the alt text on its images, whether the image is in an og:image tag and in structured data, whether it’s served in HTML or lazy-loaded by JavaScript, and — by fetching a sample of the actual image bytes — whether any carry C2PA/Content-Credentials or IPTC AI-disclosure metadata.
1,082 sites were reachable (ecommerce 219, service 467, personal 396). Every figure below is on those; proportions carry 95% Wilson intervals.
The websites built on images describe them the worst
The headline is an irony. The share of sites where most images carry no usable alt text — missing, empty, or a generic word like “image” or “logo” — climbs steadily from stores to service sites to the portfolios whose entire pitch is visual.
A photographer’s gallery with no alt text is invisible twice over: a multimodal model has to guess what’s in each frame, and a screen-reader user gets nothing. The people with the most to show are telling the machine the least about it.
Whether the image is in structured data depends on who built the site
An answer engine or shopping agent pulls the image URL from structured data, not a CSS background. Coverage splits sharply by cohort — service sites (often on templates that emit LocalBusiness/logo schema) lead; portfolios and stores trail.
The og:image tag — the single most-used machine-readable image signal, and what most AI/social previews grab — is present on 73.0% of personal, 66.0% of service, and 46.1% of reachable ecommerce sites. Even that basic tag is missing on a quarter-to-half of sites.
Almost nobody labels anything — and you can’t detect it either
You raised the obvious 2026 question: how many of these images are AI-generated? The honest answer is that you can’t reliably tell, and the metadata that’s supposed to tell you is used by almost no one.
Only ~1% of sites in every cohort (ecommerce 0.9%, service 1.3%, personal 0.6%) served any image carrying C2PA/Content-Credentials or formal IPTC AI-disclosure metadata. The provenance standards that regulators (EU AI Act) and platforms (Adobe, Google, Meta) are promoting have effectively zero real-world adoption.
We also tried to detect AI-generated images by scanning metadata for generator names — and threw the result out, because it was garbage. The most common “hit” was the string imagen, which is just the Spanish and Portuguese word for “image,” not Google’s Imagen model. That’s the real finding: naive AI-image detection produces false positives you shouldn’t publish, and the one trustworthy signal — provenance metadata — is missing from ~99% of the web’s images.
The ecommerce access gate
One more wall showed up before we could even read the images. Of 1,200 Shopify storefronts sampled, only 18% returned content to our fetch — 54% actively blocked it, even with a browser user-agent (Shopify’s bot management fingerprints more than the user-agent string). So the ecommerce cohort is a smaller, self-selected 219 sites, and we flag it as such. Service and personal sites, by contrast, were reachable ~88% of the time.
What to do about it
All of this is cheap to fix, and every fix pays off twice — for AI and for accessibility:
1. Write real alt text on meaningful images. “Blue mid-century sofa in a sunlit loft,” not “image” or “IMG_2043.”
2. Put the image in structured data — Product.image for products, ImageObject/logo elsewhere — and add an og:image tag.
3. Serve the image URL in HTML, not only via JavaScript lazy-loading, so a non-rendering crawler can find it.
4. If you publish AI-generated images, disclose them with Content Credentials — you’ll be in the ~1% that does, ahead of the labeling mandates.
Frequently asked questions
Can AI actually read the images on a website?
Only through the signals around them: alt text, og:image, the image URL in structured data, and the image bytes. A decorative image with no alt, not in schema, lazy-loaded by JavaScript is nearly invisible to a non-rendering AI crawler.
Which sites were worst?
Personal/portfolio sites — photographers, designers, artists. On 43.9% the majority of images have no usable alt text, vs 29.3% of service sites and 21.0% of ecommerce. The sites built on images describe them worst.
How many websites disclose AI-generated images?
About 1% in every cohort carry any C2PA/Content-Credentials or IPTC AI-disclosure metadata. Adoption of the provenance standards is effectively zero.
Did you detect which images are AI-generated?
We tried and discarded it as unreliable — the top false positive was 'imagen' (Spanish for 'image'). You can't reliably detect AI images from metadata, and the provenance metadata that would tell you is used by almost no one.
What should I do?
Write descriptive alt text, expose the image in structured data + og:image, serve image URLs in HTML (not JS-only), and disclose AI-generated images with Content Credentials.
Method & limits
Three cohorts: ecommerce = 1,200 Shopify custom-domain storefronts; service = 535 harvested local-business sites across five verticals and many US metros; personal = 446 harvested portfolio/personal-brand sites. Crawled 16 September 2026 with a static, non-JavaScript-rendering fetch. Per site we parsed <img> tags (real src/srcset vs JS-lazy attributes, and alt quality), detected og:image and JSON-LD image references, and byte-scanned up to five sampled images for C2PA and IPTC provenance markers (validated against the official C2PA test fixtures). Metrics are computed on the 1,082 reachable sites; proportions carry Wilson intervals.
Known limits, stated plainly: a static fetch is a conservative floor for image counts (images inside <picture>/<source> or injected by uncommon lazy-loaders are undercounted) — which is the point, since non-rendering AI crawlers face the same floor. The ecommerce cohort is a small, self-selected 219 sites because 54% of Shopify storefronts blocked our fetch and ~25% were dead on this run; its figures are directional. The service and personal frames are search-harvested rather than censuses and skew toward sites visible enough to surface in results. Provenance is measured on up to five sampled images per site, so site-level rates are lower bounds. Naive AI-generated-image detection was attempted and discarded (false positives). Raw per-site data and the crawler ship with the study.
We run this readiness check against a real site, then fix what an AI can’t read — images included.
AI Search Visibility (GEO)