Why the preview is the hard case
A link preview is built by the network, not by you. It fetches your page, reads the tags described in Open Graph tags in WordPress, downloads the image and shows a small card. Whatever you wrote in the caption on the page does not travel with it. So a generated illustration can reach thousands of feeds without a word of context, and the first time a reader learns it was generated may be in the replies.
The fix is to make sure the facts about the image are present in the three places a network or a curious reader can look: the file, the page and the alt text. The provenance and authorship section covers the wider picture; this guide is the practical checklist.
Label the image in four steps
Caption it on the page
Put a plain caption under the image: Illustration generated with an image model and edited by the author. Keep it short and specific about what was generated.
Write alt text that describes, not disclaims
Alt text is for people who cannot see the image. Describe what it shows first; add that it is an illustration if that matters to understanding. Alt text is not the place for a policy statement.
Keep the digital source type in the file
The IPTC digital source type vocabulary has values for media created by trained algorithms and for composites that include them. Tools that write it store it in the file's metadata, and platforms that read it can label the upload. See where image metadata gets stripped before you rely on it.
Check the card after publishing
Run the page through each network's preview tool and look at the card at phone size. The guide to Open Graph image size explains the crop.
What each signal reaches
| Signal | Who sees it | Survives a share? |
|---|---|---|
| Caption on the page | Readers who open the link | No, the card does not include it |
| Alt text | Screen reader users on the page; some networks copy it | Sometimes |
| Digital source type in the file | Platforms that read image metadata | Only if the file is not re-encoded |
| Content Credentials | Tools that verify C2PA manifests | Rarely after upload |
| Platform label | Everyone on that network | Set by the platform |
Meta has said it labels images when industry-standard indicators such as C2PA and IPTC metadata are present, and asks users to disclose realistic generated media themselves. That makes the file's metadata the one signal you control that can reach a feed. It is fragile, which is why the caption still matters.
An AI generated image link preview: a worked example
Take a post with a generated illustration, a caption, a descriptive alt text, the source type set in the file and a description that ends with the line: Illustration generated with an image model. In an invented worked example the post is shared to four networks and each card is inspected at phone width. Three of the four cards show the description line, so card readers on those networks see the disclosure; the fourth network shows only a title and image, so on that network the title is the only text that travels. None of the four preview images carries the source type field after the network re-encoded it. Three cards crop the 1200 by 630 image to a square, which would have hidden a badge printed in the lower corner.
A second post in the example carries the same image without the description line. Its four cards show a title and an illustration with no context at all, and the first reply on one network asks whether the picture is real. The difference between the two posts is one sentence in the description.
Pitfalls
- Printing a label onto the image. Square crops on several networks remove the corners where such labels are placed.
- Writing the disclosure into the alt text instead of a caption. Screen reader users hear it every time; sighted card readers never see it.
- Setting the source type field in the original and letting an optimiser strip it from the file that og:image points at.
- Using a stock photograph as og:image for a post whose body has generated images, so the card implies a photograph.
- Testing on desktop only. The crop and the visible description length differ at phone width.
Check it worked
After publishing, share a fresh address on each network from a logged-out or secondary account, look at the card on a phone, and download the preview image the network's debugger shows. Log one row per network, as below with example entries. Re-run the check after theme or plugin changes; the share preview integrity routine gives the full procedure.
| Page | Network | Time (UTC) | Preview showed | File carried |
|---|---|---|---|---|
| /2026/09/harvest-illustration/ | 2026-09-17 08:10 | Description ends with disclosure line; square crop | No source type field | |
| /2026/09/harvest-illustration/ | 2026-09-17 08:18 | Description trimmed but line visible | No source type field | |
| /2026/09/harvest-illustration/ | Mastodon | 2026-09-17 08:25 | Full description, 16 by 9 crop | No source type field |
| /2026/09/harvest-illustration/ | X | 2026-09-17 08:33 | Large image card, no description shown | No source type field |
What not to put on the card
A badge or watermark printed into the preview image looks like a solution and works badly. It eats the safe zone, gets cropped on some networks, and reads as a warning without explaining anything. If a label matters, make it part of the page title or the first sentence of the description, which networks do show. Readers who care will follow the link; the guide to disclosure lines that survive a share shows wording that fits.
Sources
- IPTC NewsCodes: digital source type vocabulary
- Meta, Our Approach to Labeling AI-Generated Content and Manipulated Media (April 2024)
- The Open Graph protocol
Questions
Should I label AI images in my Open Graph preview?
Label them on the page, in the alt text and in the file metadata. The preview card is too small and too often cropped to carry a useful label.
Do networks read AI metadata in shared images?
Some do. Meta has said it uses C2PA and IPTC indicators to label images, but re-encoding during upload often removes them.
What alt text should a generated image have?
A plain description of what it shows. Mention that it is an illustration only when that affects how a reader understands it.
Is it misleading to use a generated image as a featured image?
Not if it is honest about what it is. Realistic images of real events or people without disclosure are the problem.
Do I need to disclose a generated image in the og:description?
Only when the image matters to how the post is read. When it does, the description is the one text field most networks show on the card, so a short factual line there reaches card readers.
Hannah Voss, Editor. Checks every guide against a working WordPress install and the networks' current documentation. Last reviewed September 2026.
Related reading
- Where image metadata gets stripped: WordPress, CDNs and uploadsWhere IPTC, EXIF and Content Credentials metadata is lost between your camera and a reader's feed: WordPress resizing, optimisers, CDNs and social uploads.5 min read
- AI disclosure lines that survive a shareWhere to put an AI assistance disclosure so it reaches readers who only see the share card: titles, descriptions, bylines and the first lines of a post.5 min read
- Open Graph image size: 1200 by 630 and the safe zoneThe right Open Graph image size is 1200 by 630 pixels at a 1.91:1 ratio. Minimums, file limits, formats, and a centre safe zone for square crops.6 min read
- Content Credentials on shared images: what survives a shareWhat C2PA Content Credentials are, why WordPress image sizes and social uploads often strip them, and how to keep provenance on the images you share.6 min read