When an AI engine cites a source, the buyer gets a URL and a span of text. Open any answer that describes a chart and follow the link: it resolves to the page that carried the chart, and the number the answer quoted is almost always written out in the prose around it. That one mechanic decides how images, video, and structured data earn citations. Each has to feed a retrievable text record, and teams that treat the three as independent optimization channels are funding three projects to solve one problem.
The honest version has boundaries. Google says plainly that AI Overviews and AI Mode require no additional markup or special optimization to appear, only pages that are indexed and snippet-eligible, while it documents separate user agents for image and video fetching, so the modalities are crawled apart and answered together. Ahrefs found 38% of AI Overview citations ranked in the organic top 10 across 4 million AI Overview URLs in March 2026, and Seer found 87% of SearchGPT citations matched Bing's top results. Retrieval gates every surface.
Do AI engines cite an image or a video as a separate source?
Sometimes, and far less often than the multimodal pitch implies. In the engines that publish their output format, a citation is a URL paired with the text it supports: OpenAI's web search documentation describes a response with two main parts, the message and the URL citations attached to it. A video cited in an answer is usually a page on a video domain. An image cited in an answer is usually the page that hosted it.
That does not make the modalities irrelevant. Google crawls media with distinct agents and indexes it into distinct surfaces, and ChatGPT search resolves a question by issuing one or more targeted queries before it writes anything. What comes back is text. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited in the final answer, and Vercel's crawler research with MERJ documented GPTBot and its peers fetching raw HTML without executing JavaScript.
The question worth asking, then, is whether the claim your image or video carries exists anywhere the retrieval step can read it. A chart whose only copy of the number lives in pixels has published that number to nobody.
What does structured data actually do for AI retrieval?
It makes pages eligible for search features and tells parsers which entities a page describes. That is the documented job. Google's introduction to structured data presents markup as a standardized way to describe page content, the supported-types gallery lists which features each type can qualify for, and the schema.org getting-started guide covers the vocabulary itself.
What none of that documentation says is that markup earns AI citations. Google's guidance on its AI features states there are no additional requirements and no new machine-readable files or markup needed, and describes a query fan-out across sub-questions instead. A percentage lift attributed to adding schema is a correlation, and the mechanism behind it stays undocumented.
Schema still earns a day of work, for reasons that have nothing to do with citations. It disambiguates your entity for parsers, and it qualifies pages for the features the supported-types gallery lists. Finish it once, then put the rest of the budget on whether your claims are retrievable at all.
Where markup does earn its keep is provenance. W3C's provenance ontology models Entity, Activity, and Agent, which is the shape of any honest statement about who produced a claim, from what, and when. Author, publisher, date, and method are the properties that help a parser attribute a number to you rather than to whoever republished it.
How do you make an image retrievable when the engine reads text?
Write the number out. The shortest path from a chart to a citation is a sentence in the body that states what the chart shows, on the same page the chart sits on. Alt text, caption, filename, and the paragraph around the figure are all text a crawler reads; the rendered pixels are not.
Two mechanics decide whether that text survives. The figure and its caption have to sit in the HTML the crawler receives, which rules out charts injected after page load: rendering sits in a separate stage of the indexing process even at Google, and Cloudflare's agent-readiness scan of the 200,000 most visited domains found few sites have adopted the standards built for agents, with 3.9% serving Markdown on request. The description also has to match the image: Google's image guidance asks for alt text that is useful and in context of the page, and warns that stuffing it with keywords may get a site seen as spam.
A chart is a delivery format for a statistic, and the statistic is the part that gets extracted. In the GEO study (Aggarwal et al., KDD 2024), benchmark pages carrying statistics, quotations, and citations gained roughly 30 to 40% in citation visibility, and each of those signals is text. Write the figure's headline number into the caption and into the paragraph beside it, and the chart becomes legible to the step that decides citations.
How do you make video count as evidence instead of decoration?
Publish the transcript as a page you control, structured the way you would structure an article. A video citation, in practice, is a citation of a text record about the video: the transcript, the description, the chapter titles, and the page hosting them.
Section discipline matters more here than anywhere else, because extraction works on passages, not pages. One 2026 benchmark built its gold evidence passages at an average of 188 tokens, with a 56.26-token standard deviation, across 1,186 multi-hop chemistry questions. Passage-sized units are a general pattern in retrieval, but that number describes how one chemistry dataset was built, not how engines split a page or a length to write to. Cut the transcript into named sections that each stand on their own, with the buyer question in the heading and the answer in the first two sentences.
Then check that the transcript page is retrievable on its own terms. Cloudflare's agentic-internet bot report puts AI training at 52% of crawler requests, and OpenAI documents four relevant user agents doing different jobs, so a page closed to one is not closed to all. Verify which agents your transcript pages actually admit before you produce more of them.
How much of this is worth funding?
Fund retrieval and text first, and treat format production as unproven until it earns its place. C-SEO Bench (NeurIPS 2025) is the one benchmark that put conversational-SEO rewrite tactics head to head, and it found most of them largely ineffective and often harmful to citation rank across two tasks and six domains. It tested rewrites of page text rather than formats, and no published benchmark of that kind covers images or video.
Say the gaps out loud. The research does not establish that engines assign independent citation credit to an image or a video file, that adding schema raises citation probability by any stated percentage, or that a fixed ratio of formats produces a fixed ratio of answers. Those figures circulate anyway. Ask for the primary source before the number enters a plan.
Two more constraints bound what you can conclude. Measurement is noisy: a 2026 variance-components study measured how much of the gap between runs comes from the engine itself, and the share is large enough that a few runs cannot separate a real change from churn. Work on quantifying uncertainty in AI visibility reaches the same place through confidence intervals. Much of the answer is also off your domain: Profound found more than 4 in 10 AI citations point to sources brands do not control.
What survives contact with that evidence is per-surface measurement. The IAB's AI-era visibility guidance separates presence, prominence, portrayal, and persuasion, and sorts the work for brands into four groups. Judge each surface against the question it was built to answer, and drop any blended number that hides which one moved.
How does Trovance treat each surface as its own evidence record?
Trovance starts from the tracked questions your buyers actually ask, runs them every week across AI engines, and preserves every answer run as an answer snapshot with its full source list. Because the snapshot keeps the cited URLs, you can observe which surface carried the answer: your article, your transcript page, a third-party review, or a video on a channel you do not own.
That record is what makes per-surface work decidable. Share of answer shows which buyer questions you appear in at all, and comparing snapshots over time shows whether the transcript page you published started getting cited or whether the citation kept going to a competitor's roundup. The diagnosis lands on a specific surface, which is how a crawling problem stops being answered with a video budget.
A caption, a transcript section, and a schema property each need a claim you can defend and the evidence behind it. That inventory lives in the Brain, where each claim sits with its proof, so every surface can draw on the same record. Recommended actions name the asset the answer record says is missing: the transcript page for a question where a competitor's video keeps getting cited, or the comparison page the snapshots keep sourcing from somewhere else. Drafts use your author voice guidance and research sources, and a person reviews and approves everything before it publishes.
What Trovance will not do is promise that adding a format produces a citation. The evidence above rules that promise out for any tool: the engines are probabilistic, and much of the source set belongs to other people. What it does instead is close the loop: the next analysis cycle reruns the same questions so you can verify whether the answer moved after your change, and learn which of your surfaces is actually doing work.
What should you do this week?
Work in the order the retrieval step reads. Fetch three of your best pages with JavaScript disabled and confirm the captions and the transcript text survive the fetch. Then take every chart on those pages and write its headline number into a sentence of body text. Publish one transcript as a real page with question-shaped headings instead of a raw dump under an embed.
Then audit what you already believe. List the multimodal claims driving your roadmap and find the primary source behind each one. The claims with no source underneath them are budget you get back. Schema belongs with the cheap plumbing you finish in a day.
Measure per surface, and give the measurement enough runs to mean something. If you would rather have that record than build it by hand, start a free Trovance trial and see which surface is carrying the answers in your category today.
Make each surface retrievable
Schema markup and AI citations - what structured data does and does not do for retrieval.
How AI actually reads your content - what the parser sees before any answer is written.
Technical SEO for AI search - the fetch and render mechanics underneath every surface here.
Can AI find the right page in your PDF - the same retrieval problem in a non-HTML format.
Judge the evidence before you fund it
What actually gets cited by AI - which page traits correlate with being used as a source.
The most-quoted study in AI search, quoted correctly - what the GEO benchmark really tested.
AI agents need evidence, not more content - why more formats is not more proof.
Measuring AI search visibility without one score - how to read each surface separately.
