Multimodal Readiness: Images, Alt Text, and Diagrams that Models Can Cite

Multimodal Readiness: Images, Alt Text, and Diagrams That Models Can Cite

By Lukas Richter, Senior Content Automation Strategist

A diagram can explain in seconds what a page of text would take minutes to describe. But if its meaning lives only inside an image file, many readers—and many search systems—may miss the point.

That is why multimodal readiness is not about adding more pictures. It is about packaging each useful visual so its meaning is clear in the image, in the page text, and in the code that connects the two.

There is one important limit: good alt text may help a system understand an image, but it does not guarantee that an AI assistant will cite the page. Understanding an image and choosing a web page as a source are different steps.

What the research does—and does not—show

Captions can help with visual questions

Some research finds that text paired with an image can help a model answer questions about it. A 2019 study found that adding paragraph-style captions to images improved performance on a visual question-answering task. The captions supplied details that were easier to match to a question than image features alone. But this was a model experiment—not a test of whether adding alt text to a live web page increases AI citations. ACL Anthology paper (aclanthology.org)

Charts and diagrams also remain hard for models to read precisely. In one chart-question benchmark, researchers found weaknesses in tasks such as finding relationships between values. Their results improved with a special text prompt, which points to the value of clear language—but does not prove that any one page format will earn more citations. ChartInsights benchmark (arxiv.org)

The broader lesson is practical: a visual can carry useful evidence, but models may still miss small labels, exact values, or the path between steps. OpenAI’s vision guidance, for example, warns about small text, charts, color-coded elements, and precise spatial reasoning. OpenAI vision guide (developers.openai.com)

Better alt text is not a citation switch

Search systems may use image descriptions and surrounding text to understand what a picture shows. Google says it uses alt text, computer vision, and page content together when understanding images for Search. It also recommends placing images near relevant text. That supports writing useful alt text, but it is not evidence that alt text by itself makes an AI assistant cite a page. Google Image SEO guidance (developers.google.com)

The evidence reviewed for this article does not establish that changing alt text alone raises citation rates across current AI assistants. Many systems do not explain exactly which page elements they read or why they choose one source over another. Google says its AI Search features have no special markup requirement and do not guarantee that an eligible page will be shown. Google’s AI features guidance (developers.google.com)

So treat improved visual packaging as a reasonable, testable hypothesis, not a promise. It can make a page easier to use, easier to parse, and more likely to contain text that directly answers a question. Whether an assistant then cites it is a separate result to measure.

Package each important visual as evidence

A strong visual package has four parts: the image, concise alt text, a caption, and a nearby text explanation. Each part has a different job.

1. Write alt text as a short replacement

Alt text should give readers who cannot see an image the information they need to follow the page. For a simple diagram, that may fit in one sentence. For a complex chart or flowchart, alt text should identify the visual and its main purpose, not squeeze every detail into one long description.

For example:

alt="Checkout flow with a retry path after payment failure."

For a decorative image, use empty alt text (alt="") so assistive tools can skip it. Do not leave alt text out, and do not fill it with repeated search terms. W3C’s complex-image guidance and Google’s image guidance distinguish short text alternatives from fuller descriptions. (w3.org)

2. Use a caption to state the figure’s point

A caption should help a reader understand why the image matters. A label such as “Architecture diagram” is usually too vague.

A more useful caption is:

Figure 1. A failed payment sends the shopper back to payment; the cart stays open.

Use the HTML <figure> and <figcaption> elements to connect the caption to its image. A clear figure number or name also makes it easier to refer to the same visual in the article text. Google’s documentation style guide gives examples of figures paired with alt text, captions, and longer descriptions. (developers.google.com)

3. Put the full explanation in the page

A complex diagram usually says more than a short alt description can. Put the important meaning in ordinary page text too. For a flowchart, describe the steps and what happens at each branch. For a chart, include the time period, units, key values, and the trend. For a technical diagram, name the parts and explain how they connect.

W3C recommends a short description plus a fuller text description for complex images. That fuller description can be in the article itself, near the image, and available to all readers. A data table can help when exact values matter. W3C WAI: Complex Images (w3.org)

4. Keep names and labels consistent

If your diagram calls something “activation,” use that same word in the caption and article. If the figure shows a result that matters, state it in a sentence nearby: “In Figure 2, most sign-ups drop off before the first project is created.”

This is not a magic citation trick. It is a way to make the page’s evidence easier to follow and quote. Google says there is no special schema for AI Overviews or AI Mode; the useful work is keeping important content in text and making the structured data match what people can see. (developers.google.com)

Here is a simple pattern for a flowchart:

<p>The key point in the checkout flow is that a payment failure does not clear the cart.</p>

<figure id="checkout-flow">
  <img
    src="https://cdn.example.com/checkout-flow.svg"
    alt="Checkout flow with a retry path after payment failure."
    width="1200"
    height="720">
  <figcaption>
    <strong>Figure 1.</strong>
    A failed payment sends the shopper back to payment; the cart stays open.
  </figcaption>
</figure>

<p><strong>Figure 1 in words:</strong>
Cart → shipping → payment → confirmation. If payment fails,
the shopper returns to payment and keeps the cart.</p>

The short alt text identifies the diagram. The caption gives its main point. The last sentence explains the path in plain text. For a chart with exact values, add a table or list of those values too.

Make sure crawlers can fetch the image

A well-described image still needs to be available. Use a normal HTML <img> element with a real src URL. Google says it can find images in an <img> element’s src, but it does not index CSS background images in the same way. Its image guidance also supports CDN-hosted images in image sitemaps. Google Image SEO guidance (developers.google.com)

For a CDN, check the basics:

• The image URL works without a login or an expiring access token.
• The page’s crawler permissions and the CDN’s security rules both allow access.
• The image URL in the rendered page is stable and returns the correct file.
• Responsive images include a fallback src.
• Important text is not hidden in an image when it could be real page text.
• Below-the-fold images load when they become visible—not only after a person clicks or scrolls in a way a crawler cannot reproduce.

Google’s lazy-loading guidance says image URLs should appear in the rendered HTML and load when they come into view. It also advises against lazy-loading images that are already visible when the page opens. Google’s lazy-loading guidance (developers.google.com)

Check permissions by platform, too. OpenAI documents OAI-SearchBot as its search crawler and says it is separate from GPTBot, which relates to training data. Allowing one does not require allowing the other. This matters if you want search discovery but have a different preference about training use. OpenAI crawler documentation (developers.openai.com)

Test whether your pages get cited for visual questions

Do not rely on a few casual chatbot prompts. Run two tests, because model comprehension and web citation are not the same thing.

Test 1: Does the model understand the visual?

Choose a set of charts and diagrams. Write questions whose answers depend on them, such as:

• Which step follows payment?
• What happens when payment fails?
• Which month had the highest conversion rate?
• What does the diagram say is the main reason users drop off?

Then compare model answers under controlled input conditions:

1. The image alone.
2. The image plus its alt text and caption.
3. The image plus those annotations and the page’s text summary or data table.

Keep the model, question, and image the same. Score answers for correctness, including exact values and relationships. This helps show whether added text helps the model answer your questions. It does not show whether a web search assistant will find or cite the page.

Test 2: Does a search assistant cite the page?

For a small pilot, pair similar diagram-heavy pages. Add the full visual package to one group; leave the comparison group at its current accessible baseline. Do not remove essential alt text to create a control group. Prepare two sets of questions: ones that require reading a figure and ordinary questions that the page’s text already answers.

Run the same prompts in the search assistants you care about. Record the date, product, question, answer, and cited page. Repeat the checks over time, because results can change. Wait until your crawler or search tools show that the updated pages have been recrawled. Google notes that recrawling can take days or longer. (developers.google.com)

Track four outcomes:

• Citation rate: How often does the answer cite the page?
• Answer accuracy: Did the assistant get the visual fact right?
• Citation support: Does the cited page actually support the answer?
• Business result: Did the visit, referral, or lead help the business?

Compare visual-question results with the ordinary-question control set. A simple before-and-after comparison can be misleading: the assistant, its search index, and the wider web may all change at once. Matched pages and repeated checks help, but a small test should still be reported as a pilot, not a universal rule.

Also be precise about what the result means. If a page gets cited more often after you add a text summary, the test shows that the whole package helped. It does not prove the assistant read the image itself. Citation dashboards can show which URLs appear in AI answers, but not necessarily which part of a page was read. Bing’s AI Performance report, for example, reports cited pages and query information. Bing Webmaster Tools: AI Performance (blogs.bing.com)

Build the work into your content system

Founders should not have to rewrite a diagram by hand every time they publish an article. But automation should not invent what the picture means.

A simple workflow is:

1. The creator provides the source data or diagram, plus the main point they want the reader to take away.
2. A content workflow drafts the alt text, caption, and plain-text summary from those source materials.
3. A person checks that the description is accurate and that labels, values, and routes match the visual.
4. The page publishes the image and its text together.
5. A repeatable check confirms that readers and crawlers can access both.

The creator’s point is the part that should not be automated away. For example, a founder may want to show not just that sign-ups fall, but that the team chose to fix first-project setup before buying more traffic. That is an opinion and a product lesson, not a captioning task.

In an automated content engine, the Creator, Audience, and Taste inputs can preserve that point across articles. AutoPod uses those inputs to guide each article, while the workflow handles repeatable tasks like drafting captions and checking page structure. The model can package the insight. It should not make up the insight.

Conclusion

The strongest case for descriptive alt text, captions, figure labels, and nearby summaries is not a promised SEO lift. It is that they make a visual easier to understand, easier to access, and easier to connect to the page’s argument.

That may also help an assistant answer a visual question and cite the page—but the lift is not guaranteed. Package diagrams as evidence, make the image fetchable, and measure citations separately from comprehension. That is a simple system with value for readers first and a testable upside for AI discovery.