Do Your Images Appear in AI Answers? What Is Known, and What Is Guesswork

There is a genuine question here and a great deal of noise around it. The question: when an AI answer surface displays an image, how did that image get chosen, and can you influence it? The noise: a growing body of advice that describes ordinary image SEO in new vocabulary and presents it as a technique for a new channel.

This is what can be said with confidence as of August 2026, what is reasonable inference, and what is being asserted without support.

What is actually happening

Several answer surfaces display images.

Google's AI Overviews can show images alongside a generated response, drawn from pages the system referenced. Assistant products with search capability — Perplexity most visibly, and others to varying degrees — display images from the sources they cite. The behaviour is inconsistent across query types and changes frequently, sometimes between one week and the next.

Two things follow from that inconsistency, and both matter more than any tactic.

There is no opt-in. No schema type requests inclusion, no meta tag signals availability, no console setting exists. Eligibility comes from being indexed on a page the system drew on.

There is no report. Search Console does not separate AI Overview appearances into their own dimension, and third-party assistants publish nothing equivalent. Anyone presenting a chart of their "AI image visibility" is presenting a proxy or an estimate.

The three-step chain, and where you have leverage

For an image to appear in an AI answer, three things must happen in order.

1. The page is selected as a source. This is a text and authority problem, not an image problem. If the page is not among those the system draws on, its images are not candidates. Nothing about the image changes this.

2. The image is discoverable and indexed. It must be fetchable, present in the rendered HTML, not blocked, and of sufficient quality to be useful. This is exactly the ordinary indexing pipeline, and it is where most sites actually fail — an image absent from Google Images is absent from everything downstream of it. The ordered list of reasons images do not get indexed is the highest-yield diagnostic available here.

3. The image is chosen as relevant to the answer. The system decides which of a page's images illustrates the point it is making. This is the only step where image-specific work plausibly helps, and it is the step nobody outside the providers can describe with confidence.

Notice the shape: step one is a content problem, step two is a technical problem, and only step three is the interesting new question. Most published advice conflates the three, which is how "optimising for AI answers" ends up meaning "do image SEO".

What plausibly influences step three

No provider documents a selection mechanism, so what follows is reasoned inference from how these systems are constructed, and should be held loosely.

One clear subject per image. A system trying to illustrate a specific claim needs an image that unambiguously depicts a specific thing. A composite showing five concepts at once is harder to attach to any single statement. This is the same discipline that helps in visual search, where matching depends on a clearly identifiable subject rather than on metadata.

Text that states what the image shows. Captions, adjacent copy, headings and alt text collectively establish the link between a visual and a claim. If the only thing near an image is a paragraph about something else, no consumer of that page — human or otherwise — can connect them.

Proximity to the relevant passage. An image placed beside the paragraph it illustrates is easier to associate with that paragraph than one placed at the top of an article covering nine topics.

Accurate alt text. It is part of the parsed HTML and available to anything reading the page, so assume it is read. Write it as an accurate description of the image — which is the correct behaviour for accessibility reasons that require no speculation about weighting.

Every item on that list was already correct practice. That is not a coincidence and it is not a disappointment; it means the work is already justified.

The control that genuinely exists: crawler access

The one lever with a documented, predictable effect is the one people discuss least: whether you allow AI-specific crawlers to fetch your site at all.

The major operators publish named user agents and honour robots.txt directives aimed at them. Blocking a given agent generally removes you from the surfaces that agent feeds — including whatever referral traffic those surfaces would have sent.

Choice	Effect	Suits
Allow all AI crawlers	Eligible for citation and image display everywhere	Most businesses seeking visibility
Block training crawlers, allow search crawlers	Content excluded from model training, still eligible for cited answers	Publishers with licensing concerns
Block everything	No AI surface presence at all	Sites with a specific legal or contractual reason

Two practical notes. First, the distinction between training crawlers and search crawlers is one that operators define and can change; it is a policy boundary, not a technical guarantee. Second, robots files get copied between sites without review — it is genuinely common to find a site blocking the search crawler of the exact assistant its owner is trying to appear in. Read your own file rather than assuming.

Measurement, honestly

There is no clean measurement, so use proxies and be explicit that they are proxies.

Search Console, Search type: Image. The best available signal for whether your images are being surfaced anywhere. It does not isolate AI surfaces, but a site with negligible image impressions is not appearing in them either.

Referral traffic from assistant domains. Visits arriving from assistant products appear in analytics as ordinary referrals. This measures the answer's effect, not the image's, but it establishes whether you are being cited at all.

Manual sampling. Pick ten questions a customer would genuinely ask, run them across the surfaces you care about once a month, and record whether you appear and whether an image is shown. It is unglamorous and it is the only direct observation available.

Not a proxy: rankings for the equivalent text query. The overlap between what ranks and what gets cited is real but far from complete, and treating one as a measure of the other will mislead you.

The kinds of images that get used

An observation rather than a finding, drawn from watching these surfaces rather than from any provider documentation: the images that appear alongside generated answers skew heavily towards a few types.

Images that carry information. Diagrams, comparison charts, labelled illustrations, step sequences. An answer explaining a process is well served by an image of that process, and poorly served by a photograph of somebody looking thoughtful.

Images of specific, identifiable objects. A particular part, tool, product or component. These pair naturally with answers that name the thing.

Screenshots of interfaces. For any answer about software, a screenshot of the relevant screen is the most useful visual available and there is rarely much competition for it.

What appears least often is exactly what fills most business websites: generic contextual imagery. A photograph that could illustrate any of four hundred articles is not a strong candidate to illustrate one specific claim.

The practical implication is a content decision rather than a technical one. If you want a realistic chance of an image being selected, publish the diagram you have been meaning to draw, or the screenshot, or the labelled photograph of the actual object. That advice is unaffected by whatever the providers do next, because those images are also better for the human reader — which remains the only reliable basis for a decision in an area this poorly documented.

What not to do

A few tactics circulating that are unsupported and carry cost.

Adding structured data hoping to force image inclusion. Image markup governs eligibility for specific, documented features — product images, article thumbnails, licence badges. It is not a general request mechanism, and the markup that Google actually reads does not include an AI-answer property.

Stuffing alt text with question phrasings. This degrades the field for the users it exists to serve, and there is no evidence it helps anything.

Publishing more images per page. Volume does not increase the chance of selection and does measurably slow the page.

Reworking a content strategy around it. The pages that get cited are the pages that answer questions well. That is the same investment as before, and the images are a downstream beneficiary of it.

A defensible position to hold

If someone asks what your policy is on images in AI answers, this is a position that will still be defensible in a year:

1. Every image is crawlable, indexed and present in the rendered HTML.
2. Each image shows one clear subject and sits next to text that explains it.
3. Alt text is accurate, written for people, and complete on meaningful images.
4. Crawler access is a deliberate decision recorded somewhere, not an inherited robots file.
5. Image impressions and assistant referrals are tracked as proxies, with the limitation stated.
6. No budget line exists for AI image optimisation specifically, because no technique justifies one yet.

That last point is the one worth defending hardest. The correct response to an under-documented surface is to do the well-understood work properly and stay ready to move, not to buy a tactic nobody can verify.

SEOpix generates images with descriptive filenames, accurate metadata and a single clear subject — which is the part of this that is knowable today. See the plans, or start on the free tier.