Every SEO team is asking the same question right now: how do Perplexity, ChatGPT, Gemini, and Claude find images, and how do you get yours cited? The honest answer is that we know less than the confident blog posts suggest. But “less than we’d like” is not “nothing,” and the things we do know point to a reassuringly familiar set of actions.
This article separates what is documented from what is speculation, and ends with the handful of moves actually worth making. It is the AI-search layer of our complete image SEO guide.
The state of AI image search in 2026
Two patterns are clear from crawler traffic data. First, the majority of AI crawler activity is for training rather than live retrieval — the bots are building models more than answering this minute’s question. Second, when AI answers do cite sources, they lean heavily on a small set of domains, and a brand’s own domain features prominently among its own citations. The implication: being crawlable and being clearly attributable both matter.
What no engine has published is the part everyone wants — exactly how it decides which image to show in an answer. Anyone claiming a definitive method is guessing. So the sensible strategy is not to chase an undocumented ranking system; it is to make your images maximally understandable, on the assumption that understandable images are easier to choose.
How the major engines differ
- Perplexity — the most aggressive crawler, and the most controversial: in 2025 Cloudflare documented it using undeclared crawlers to reach pages that had asked not to be crawled. Whatever your stance, assume Perplexity is thorough.
- ChatGPT — backed by Bing’s index for live results, so classic SEO that helps you in Bing tends to help you here.
- Gemini — draws a large share of citations from a brand’s own domain, which rewards having clear, well-structured pages of your own.
- Claude — a well-behaved crawler that declares itself and honours
robots.txt, which makes your robots file an effective control for it.
The bots and their user-agents
You control AI crawler access the same way you always have — through robots.txt, by user-agent. The image-relevant names to know in 2026 include GPTBot, ClaudeBot, PerplexityBot, and ImagesiftBot, the last of which is a dedicated image-training crawler. Naming them in your robots file lets you allow or disallow each deliberately.
Why server-rendered images matter more now
Here is the single most practical fact in this article: most AI crawlers have limited or no JavaScript rendering. An image injected into the page by a script after load may simply not exist as far as these bots are concerned. Images that are present in the server-rendered HTML — a real <img> with a real src and real alt text — are the ones that get seen. This is the same reason server-rendered markup helps traditional SEO, now with higher stakes.
Do this now. Open one of your gallery pages, then view the page source (not the inspector — the raw source, view-source: in the URL). Search for one of your image filenames. If it is not in the raw HTML, your images are being injected by JavaScript and may be invisible to AI crawlers. That is a strong reason to use a gallery that outputs real <img> markup server-side.
Schema as the discoverability layer
If an AI engine is trying to understand and attribute an image, structured data is the cleanest signal you can give it. An ImageObject with a clear caption and a named creator tells the engine what the image is and who it belongs to — exactly the information needed to cite it correctly. Schema was always a discoverability layer for Google; for AI search it is also an attribution layer.
The robots.txt decision
This is the one genuinely AI-specific choice, and it is a real fork. Allow the AI crawlers and you become eligible to be cited in AI answers and included in training data. Block them and you keep your images out of both. There is no universally correct answer — a photographer protecting a portfolio may choose differently from a marketer chasing reach. The mistake is not choosing; it is leaving the default in place without having decided.
Tip. You do not have to treat all AI bots the same. You can allow the well-behaved, declared crawlers and disallow the ones whose behaviour you are less comfortable with — robots.txt lets you set rules per user-agent.
What we genuinely don’t know
It is worth being plain about the gaps, because the field is full of confident claims that outrun the evidence. No AI engine publishes how it picks which image to cite. We do not have reliable image-citation analytics the way we have Google Images data. And the engines change behaviour frequently and without notice. Treat anyone selling certainty here with caution, and treat your own approach as a hypothesis you revisit, not a settled strategy.
Working the robots.txt decision through, by site type
Because this is the one genuinely AI-specific choice, it is worth working through rather than hand-waving. The decision is not binary across all bots — robots.txt lets you set rules per user-agent — so the real question is which bots, for which goal. A few honest positions by site type.
A marketer or business chasing reach generally wants in. Being cited in AI answers is distribution, and in a world where AI search is taking a slice of traditional clicks, being a source the engines can quote is closer to an opportunity than a threat. Allow the well-behaved crawlers and accept that your images may appear in answers — that is the point.
A photographer protecting a portfolio faces a genuine tension. The same images you want discoverable are images you may not want absorbed into training data or reproduced in an answer without attribution. A reasonable middle path is to allow the crawlers that declare themselves and pass attribution (so you are cited, with a link) while disallowing the dedicated image-training crawlers if being in a training set concerns you. There is no universally right answer here; it depends on how you weigh reach against control of your work.
A shop usually wants product images findable everywhere a buyer might look, AI answers included — the calculus is the same as for any other discovery channel. The caution is less about blocking and more about making sure the images that do get crawled carry accurate descriptions and prices in real text, so an AI answer represents your product correctly rather than guessing.
The thread through all three: decide on purpose. The failure mode is not “allowed” or “blocked” — both are defensible — it is leaving the default in place because you never looked, and discovering later that you are either invisible to or fully absorbed by systems you never made a choice about.
Do this now. Open yourdomain.com/robots.txt in a browser and read what is currently there. Most sites have never added rules for the AI bots, which means the default applies by accident. Decide, for your site type, whether you want GPTBot, ClaudeBot, PerplexityBot, and ImagesiftBot allowed or disallowed, and set those rules deliberately. Even “allow all of them” is a fine answer — the point is that it is now a decision.
Five things worth doing this quarter
- Make sure your images are server-rendered — real
<img>tags in the raw HTML, not script-injected. - Write genuine alt text and captions — the text these engines read to understand an image.
- Add ImageObject schema where it earns its place — caption and creator give attributable signals.
- Make your robots.txt decision deliberately — per user-agent, on purpose.
- Keep doing classic image SEO — Google’s own guidance says good image SEO is AI optimisation.
None of these is exotic. They are the same fundamentals that serve Google and accessibility, which is the quietly reassuring conclusion: you do not need a separate AI-image strategy. You need the basics done well, server-rendered, and a deliberate robots.txt decision on top.
Key takeaways
- Most AI crawlers can’t run JavaScript well — server-rendered images are essential.
- Schema with caption and creator is the attribution layer for AI citation.
- Your robots.txt is a real, per-user-agent decision — make it on purpose.
- There is no separate AI checklist; good image SEO is AI optimisation.
Can AI search engines see images loaded by JavaScript?
Often not. Most AI crawlers have limited or no JavaScript rendering, so images injected after load may be invisible to them. Use real server-rendered <img> markup so your images appear in the raw HTML.
Should I block AI crawlers in robots.txt?
It depends on your goals. Allowing them makes you eligible for AI citations and training inclusion; blocking keeps your images out of both. You can set rules per user-agent — allowing well-behaved bots and blocking others. The key is to decide deliberately.
Is there a separate strategy for ranking images in AI search?
No. The same alt text, captions, schema, and server-rendered markup that help Google also help AI engines. Google’s own guidance states that good image SEO is AI optimisation. The only AI-specific decision is your robots.txt.