Your images reach AI answers as text

Multimodal GEO is the practice of making the images on a page legible to AI answer engines, and for most engines today that means optimizing the text attached to an image rather than the image itself.

That distinction matters because the two halves of the pipeline behave differently. On the input side, engines increasingly accept a photo as the query. On the retrieval side, the thing that decides whether your page enters the candidate set is still overwhelmingly text. A diagram that carries the whole argument of your page can be, from the retriever's point of view, an empty box.

#Half the crawlers never see your alt text

Writesonic ran a controlled test in March 2026: a page seeded with 62 distinct content injection methods, each tagged with an invented marker word, then fetched by six AI assistants (ChatGPT, Claude, Gemini, DeepSeek, Grok, Copilot). Image alt text came back from only three of the six. Claude, Gemini and Copilot captured it; ChatGPT, DeepSeek and Grok did not, according to the study writeup.

The same test found content loaded through IntersectionObserver scoring zero out of six, and no crawler scrolling the page, per Writesonic. If your images lazy-load their captions, or if the caption is painted in after hydration, it is not a question of which engines read alt text. None of them get that far.

The baseline is worse than the crawler split suggests, because the attribute is often missing in the first place. The WebAIM Million 2026 report, an automated scan of the top 1,000,000 home pages, found 16.2% of home page images had missing alternative text, averaging 10.8 such images per page, and 53.1% of home pages carried at least one. The good news in the same report is direction: that 16.2% was 18.5% a year earlier.

So for a large share of the web, the image channel is not underoptimized. It is absent.

#Moving the caption into the body is a real lever, and a small one

The first published attempt to treat this as a GEO problem rather than an accessibility one is Caption Injection, which extracts a caption for each image and injects it into the page's text so that visual semantics enter the text retrieval path. The authors evaluated it on MRAMG, a multimodal retrieval benchmark of 4,800 query-content pairs, and reported it beating text-only GEO baselines under G-Eval in arXiv 2511.04080.

The absolute numbers are the interesting part. Averaged visibility improvement came to 1.85% in the unimodal setting and 1.09% in the multimodal setting, against a ceiling of 0.64% for the traditional SEO-style optimizations they compared with, per the paper. Real, measurable, and roughly the size of a rounding error in a quarterly report.

Per-dataset results in the same paper are more useful than the average: Recipe 1.85%, Wiki 1.10%, Arxiv 1.09%, WIT 0.68%, Web 0.18%, and Manual at -9.12%. The one large number is the negative one. Manual pages in that benchmark are long, with an average text length above 6,000 characters, and stuffing generated caption text into an already dense technical document made the page worse, not better.

There is a second finding worth carrying: in the ablation, the original human-written captions outperformed both VLM-generated captions and machine-refined ones, scoring 1.18 against 1.06 and 1.11 respectively in arXiv 2511.04080. The cheapest version of this tactic is also the best one. Generating alt text in bulk with a vision model is the weakest form of the intervention the research supports.

#Where pixels do count, they are a ranking input

The pixels are not irrelevant everywhere. Where a vision-language model is doing the ranking, as in product retrieval, the image is a genuine input and can be pushed on.

A January 2026 paper introduces MGEO, which optimizes an imperceptible image perturbation and a text suffix together against a VLM ranker over Amazon product listings. Text-only manipulation moved average rank by -0.73 and image-only by -1.30, while the joint attack reached -2.25, exceeding the sum of the two unimodal effects, according to arXiv 2601.12263. A heuristic baseline managed -0.30.

Read that as a measurement of channel weight rather than as a tactic. It is an adversarial paper with no defense proposed, and the effect swung hard by product category, from -3.9 on one category to +0.4 on another in the same work. The takeaway for a legitimate site is narrow: in surfaces where a VLM ranks candidates, image content carries independent weight and interacts with the text rather than adding to it.

#The input side is not the retrieval side

It is easy to conflate the two, because the product marketing points at the input. Google describes AI Mode's visual search as multi-object reasoning over a photo followed by a fan-out, with the model "doing a dozen searches for you in the time it takes to do one," in Google's own explainer.

Note what that architecture implies. The user's image becomes a set of text queries. Your page then competes on those queries the way any page does. Your images did not enter that competition unless something textual represented them.

#What to do with this

Treat images as a gate, not a lever. Four steps, in order of return:

1. Make sure alt text exists and is in the served HTML. This is the largest available gain for most sites, given the WebAIM baseline, and it costs nothing.
2. Write the caption yourself. The ablation says human captions beat generated ones, so bulk vision-model alt text is the weak version.
3. Put the load-bearing content of a chart or diagram in body text near it, not only in the alt attribute. Alt text reaches three of six crawlers; body text reaches all of them.
4. Do not do this on long technical pages that already answer the question. That is where the benchmark measured a loss.

The honest summary of the current evidence: fixing missing alt text is a correctness fix with real upside, moving caption content into prose is worth roughly one to two percent of visibility on short and medium pages, and everything beyond that is unsettled. Anyone selling image optimization as a major GEO channel is ahead of the research.