Do Product Facts in Images Limit AI Search Visibility?

Executive summary

Yes. A fact that lives only in a picture can be missed when an AI agent reads your page. Many agents fetch the text of a page and stop there. The picture never reaches them. Some agents do look at images, but they read them with mixed results. So you can’t count on it.

The gap is bigger than most teams think. Across the top one million home pages, 16.2% of images carry no alt text at all (WebAIM, Feb 2026). Where a model does look at a chart, the best one read real-world charts right 63.0% of the time. People scored 93% (ChartMuseum, Feb 2026).

The fix is in your hands, and it’s cheap. Find the product facts that exist only in images. Write each one out as plain text, with the same units and conditions the image shows. Then check what an AI agent fetches from the page. This guide shows you how.

One page, two readers. The human sees the chart. The text-fetching AI agent gets an image slot and whatever alt text it carries. On the top one million home pages, 16.2% of images have no alt text (WebAIM Million, data from February 2026). A fact that only exists in the chart is at risk either way.

Key takeaways

• Image-only facts are a real visibility gap. An agent that extracts page text can’t read a number that exists only in pixels. Some agents also look at the image. Which ones, and how well, varies by vendor and tool. A buyer-relevant fact should never depend on it.
• Alt text is thin cover. 16.2% of images on the top one million home pages have no alt text. That is 10.8 missing per page (WebAIM Million, February 2026). Where alt text exists, it usually names the image. It rarely states the fact in it.
• Models that look at images still misread charts. On ChartMuseum’s 1,162 real-world chart questions, humans scored 93%. The best model, Gemini-2.5-Pro, scored 63.0% (Tang et al., Feb 2026). Across 14 models, accuracy fell by as much as 33.9% under realistic conditions such as a modality gap (RobustChart360, EMNLP 2026).
• When text and image disagree, the text tends to win. The same EMNLP study found that models “often over-rely on textual priors when visual and textual signals conflict”. A wrong caption can outvote a right chart.
• Restate, don’t reinterpret. The text version of an image fact keeps its units, conditions and caveats. An ambiguous visual becomes a hedged sentence. It never becomes a new product claim.
• If the human page can’t change, the AI-readable version can carry the text. SiftServe’s page-translation AI agent can run a vision pass for facts held only in images, a per-domain setting you switch on. You review and approve every draft before it ships.

This guide covers product facts that exist only inside images on your own site, as the retrieval tools and studies cited stand on 3 November 2026: specifications in graphics, comparisons in screenshots, qualifications in chart labels. It does not cover image search, alt text for accessibility compliance, or decorative imagery, and it does not claim that any AI agent is unable to interpret images. It’s written for marketers, product managers and web leads who own a product site and want AI agents to read its facts correctly.

In this guide

• Image-only facts can limit AI visibility
• What counts as an image-only product fact
• What the AI agent can extract from the page
• Which image facts need explicit text
• Keep the restatement accurate
• Check the image against the fetched text
• If the human-facing page cannot change
• FAQ
• Where to start

Image-only facts can limit AI visibility

An image-only product fact is a buyer-relevant fact that appears on a page only inside an image: a battery rating in a spec graphic, a “yes” tick in a comparison screenshot, a certification in a chart label. An AI agent that extracts the page’s text can miss it, because the fact was never in the text. Some agents also interpret images, with accuracy that varies by model, image and condition. Relying on that is a gamble with your product facts as the stake.

The direct answer to the title question is yes, with one qualifier. Image-only facts don’t make a page invisible. They make specific facts invisible, and those are usually the ones a buyer asks about. The price tier that’s only in the pricing graphic. The operating temperature that’s only in the chart axis. The “ISO 27001” that’s only on the badge.

Two numbers frame the size of the problem. On the top one million home pages, 16.2% of images carry no alt text, which works out to 10.8 missing per page (WebAIM Million, February 2026). And when a model does get to look at a chart, it’s not a safe reader. On ChartMuseum, a benchmark of 1,162 expert-written questions over real-world charts, humans scored 93% and the best model 63.0% (Tang et al., UT Austin, revised Feb 2026). Those are two different failure points: the fact never arriving, and the fact arriving wrong.

None of this promises more citations once you fix it. Citation depends on far more than one fact. What it does is close the gap between what your page says and what an AI agent can repeat about it.

What counts as an image-only product fact

A product fact is any claim a buyer would use to compare, qualify or choose. It’s image-only when the page gives it no text equivalent: no sentence, no table cell, no list item, and no alt text that states the fact itself. “Chart” as alt text doesn’t count. “Battery life: 12 hours at 20 °C” does.

Where it hides	Example	What a text-only reader gets
Specification graphic	A spec sheet exported as PNG: “IP67 · 340 g · 12 h battery”	An image slot, perhaps “spec-sheet.png”
Feature-comparison screenshot	A tick-and-cross grid of your plan vs two competitors	Nothing. Ticks have no text form
Chart axis or label	“Tested at 20 °C, 50% brightness” printed under a bar chart	Nothing. The caveat vanishes with the chart
Certification badge	ISO 27001, SOC 2, CE marks as logos	“badge” or an empty alt
Pricing graphic	Three plan cards flattened into one image by a page builder	Nothing, including the prices
Infographic	“Cuts onboarding from 14 days to 3” inside a designed figure	Nothing, unless the caption repeats it
Product photo with on-image text	A render showing “2 TB” on the drive	Whatever the alt text says

The last column is the test. Cover the image, read the page, and ask whether the fact survived. If it didn’t, it’s image-only.

Page builders make this worse than designers intend. A pricing table built in a visual editor can ship as one flattened image. A “trusted by” logo wall exported as 01.svg to 38.svg carries no customer names anywhere in the text. We see both weekly in the pages SiftServe processes, and the second is why our own extraction stage has a dedicated pass for logo walls.

What the AI agent can extract from the page

From an AI agent’s point of view, a product page holds two kinds of content, and the difference decides everything else in this guide.

Text present in the fetched page. Headings, paragraphs, table cells, list items, alt attributes, visible captions, structured data. This is what a fetch returns, and it’s what most retrieval paths pass to the model. Anthropic’s web fetch tool, for example, returns “the full text content” of a page as plain text, supports only text, HTML and PDF content types, and doesn’t run JavaScript (Anthropic developer docs, read 3 Nov 2026). Images aren’t in that list.

Facts that require interpreting pixels. The number on the chart, the tick in the grid, the words on the badge. To get these, something has to download the image and run a vision model over it. Some systems do. Google states that it “uses alt text along with computer vision algorithms and the contents of the page” to understand images (Google Search Central, read 3 Nov 2026). OpenAI’s crawler documentation describes what GPTBot, OAI-SearchBot and ChatGPT-User are each used for and says nothing about image interpretation either way (OpenAI, read 3 Nov 2026). And a vendor that does ship image reading can still scope it narrowly: Gorgias says its support AI agent’s Vision feature reads images customers send in messages but cannot “read images from your knowledge sources, including your help center articles, public URLs or uploaded documents”, and tells merchants to make sure “any important information conveyed in images is also written as text” (Gorgias help centre, read 3 Nov 2026).

Image handling differs by vendor, by tool, and by the retrieval path a given query takes. Treating every AI agent as blind would be wrong. Treating every one of them as sighted would be wrong too, and it costs more when it fails.

And sight is not accuracy. The RobustChart360 study tested 14 vision-language models on 2,505 chart instances under realistic variations: low-resolution renders, mobile layouts, a gap between what the text and the image say. Accuracy dropped by as much as 33.9% in the worst conditions, and the models “often over-rely on textual priors when visual and textual signals conflict” (Fime et al., EMNLP 2026 Industry Track). When your caption and your chart disagree, the model is inclined to believe the caption. The text you write around an image is doing more work than the image.

Which image facts need explicit text

The parent guide already covers extractable formatting in general: core answer in the first 150 words, tables for specs, FAQ for long-tail questions (Make Important Information Easy to Extract). This section is narrower. It’s about the specific claims hiding in images that need a text twin.

Start with whatever a buyer would put in a comparison spreadsheet. Four kinds of fact qualify.

1. Claims. What the product does or achieves. “Cuts onboarding from 14 days to 3.” “Works offline.” If it’s in the hero graphic, it needs to be in a sentence.
2. Values. Numbers with units. Battery hours, weight, throughput, price, seats, storage, response time. These are the facts most often locked in spec graphics, and the ones an agent is most often asked for.
3. Labels. Names and categories that let an agent match your product to a question. Plan names, model numbers, the competitor columns in a comparison grid, the customer names on a logo wall.
4. Qualifications. Certifications, compliance marks, test conditions, regional availability, warranty terms. These are the facts that decide whether you’re eligible for a shortlist at all, and they’re the ones most often shipped as badges.

What you don’t need to restate: decorative imagery, product photos with no on-image text, and anything the page already states in a sentence or table next to the image. Don’t write “the image shows our logo”. Restate facts, not pictures.

A quick way to prioritise: take your five most-visited product pages and, for each image, ask one question. If a buyer could only see the text, would they still have this fact? Every “no” goes on the list, and the list is usually shorter than you fear. Three to eight facts per page is typical in the sites we’ve sifted.

Keep the restatement accurate

The restatement’s job is to put into text exactly what the image asserts, with the same precision and the same limits. Resist improving on it.

Three rules keep it honest.

Preserve the units. “12 hours” stays “12 hours”. Not “all day”. If the chart says 11.8, the text says 11.8, even if the marketing copy elsewhere rounds it. Precision is what makes a fact quotable, and an AI engine lifting the sentence will carry your number as written.

Preserve the conditions. A chart footnote reading “tested at 20 °C, 50% brightness, Wi-Fi off” is part of the fact. The restated sentence is “Battery life is 12 hours, tested at 20 °C with the screen at 50% brightness and Wi-Fi off.” Dropping the conditions turns a test result into a promise you didn’t make.

Preserve the caveats. “Up to”, “typical”, “in our lab”, “pre-release firmware”. Keep them. If the image says “up to 2 TB”, the text says “up to 2 TB”, because the difference between “up to” and “2 TB” is a refund.

And one rule in the other direction. An ambiguous visual must not become a new product claim. A bar that looks taller than a competitor’s is not the sentence “faster than Competitor X”. If you can’t read a specific value and condition off the image, write the hedged version or nothing. “The chart compares throughput across three plans; the Pro plan is the highest of the three” is a fair restatement of an unlabelled chart. “Pro is 3× faster” is an invention.

In the image	A faithful restatement	An unfaithful one
“Up to 12 h battery*” with footnote “*20 °C, 50% brightness”	Up to 12 hours of battery life, measured at 20 °C and 50% brightness.	12-hour battery life.
Comparison grid: your plan ✓, Competitor A ✗ for “SSO”	Single sign-on is included in the Business plan. (Name competitors only if the grid does.)	The only plan with SSO.
ISO 27001 badge, no date	ISO 27001 certified. (Add the date and scope if you have them.)	Fully certified and compliant.
Unlabelled bar chart, Pro bar tallest	The chart compares the three plans; Pro shows the highest throughput of the three.	Pro is 3× faster than Starter.

Put the restatement where an agent will find it near the image: a visible caption, a sentence in the paragraph above, a row in the spec table. Alt text alone is the weakest option, because many systems truncate it and it never appears to a human reader who could catch an error.

Check the image against the fetched text

Now verify, rather than assume. The method is simple. Fetch the page the way an AI agent does, strip it to the text an agent can read, and compare that text against the facts you can see in the images.

1. Fetch the page as text. From a terminal, curl -sL https://your-site.com/product-page returns the HTML an agent fetches. Strip the tags and you have the fetched text. Our capture-stats post walks through the exact snippet, including how to measure what share of the download is readable text at all (How we measure what AI agents actually read). On our own homepage that share was 3.2% of a 382 KB download when we measured it on 31 August 2026.
2. Open the page in a browser beside it. This is the human view, images included.
3. List each image fact. Walk the images. For each buyer-relevant claim, value, label or qualification, write it down.
4. Search the fetched text for each one. Search for the fact itself, “12 hours” or “ISO 27001”, rather than the image filename. A hit in a caption, a table or a sentence passes. A hit only in alt text is a weak pass. No hit fails.
5. Fix the fails and re-fetch. The re-fetch is the point. A caption added in a page builder sometimes renders as another image.

If you’d rather not run a terminal, our readiness guide lists the same checks at the page level, with a worked example of which pages and facts to test first (How Can I Tell If Our Pages Are Ready for AI Search?). Either way, the test is the same question in two forms. Does the fact exist in the text the agent gets?

What the numbers say about trusting the image. On ChartMuseum’s 1,162 real-world chart questions, humans scored 93%, the best model 63.0% and the best open-source model 38.5% (Tang et al., revised February 2026). Across 14 models, accuracy fell by as much as 33.9% under realistic conditions such as a modality gap (RobustChart360, EMNLP 2026). And 16.2% of images on the top one million home pages carry no alt text (WebAIM Million, February 2026).

If the human-facing page cannot change

Sometimes the page can’t be touched. The pricing graphic is a marketing asset with its own approval chain. The spec sheet comes from the manufacturer. The site is on a platform where captions are images too. In that case the fix moves off the human page and onto an AI-readable version of it.

An AI-readable version is a second copy of the page, built for machine readers and served only to them, while human visitors keep the original. SiftServe builds that copy. Its page-translation AI agent rebuilds each page as semantic HTML with structured headings, plainly stated facts, FAQ pairs and Schema.org structured data, and, with the vision setting switched on for your domain, it runs a vision pass for facts contained only in images. The vision output is grounded against the page: a badge or a label is kept only when the page itself can back it, so the AI-readable version restates what your page shows and no more. Every claim traces to the original page.

Nothing ships on its own. SiftServe requires you to review and approve every draft before it goes live, which is where you catch a restatement that dropped a condition. Once approved, SiftServe’s edge worker serves the AI-readable version only to AI crawlers and agents. Human visitors continue to see the existing site, unchanged.

The full process, from checking whether your site is AI-unreadable to verifying the served copy, is in How to Make Any Website AI-Readable Without Rebuilding It. The step-by-step of how a page moves from your sitemap to an approved version is in How Does SiftServe Turn My Pages Into Approved AI-Readable Pages?.

FAQ

Can AI agents read text inside images?

Some can, some can’t, and the ones that can aren’t reliable at it. Many retrieval tools return page text only; Anthropic’s web fetch tool, for instance, supports text, HTML and PDF and returns the page as plain text (Anthropic docs, Nov 2026). Where a vision model does read a chart, the best scored 63.0% against a human 93% on real-world charts (ChartMuseum, Feb 2026). Treat image interpretation as a bonus, never as the only copy of a fact.

Is good alt text enough to make an image-only fact AI-readable?

It’s better than nothing and worse than a sentence. Alt text that states the fact (“Battery life: 12 hours at 20 °C”) does put it in the fetched text. But alt text is often truncated, rarely reviewed, and invisible to the humans who’d spot an error. On the top one million home pages, 16.2% of images have none at all (WebAIM, Feb 2026). Put the fact in a caption or a table as well.

Should I remove the images and replace them with text?

No. Keep the image for the human reader and add the text for the machine one. The two aren’t in competition. A chart with a caption that states its values serves both, and the caption also protects you when a model reads the chart wrong, since models tend to trust text over pixels when the two conflict (RobustChart360, EMNLP 2026).

What if the image fact is ambiguous or unlabelled?

Restate only what you can read off it with its conditions. An unlabelled bar chart becomes “Pro shows the highest throughput of the three plans”, not “Pro is 3× faster”. If you can’t say it precisely, say it hedged or leave it out. A restatement must never add a product claim the image doesn’t make.

Does fixing image-only facts get my pages cited more?

We can’t promise that, and nobody should. Citation depends on the question asked, the competing sources and the engine’s own selection rules. What the fix does is make sure the fact is present in the text an AI agent retrieves, which is a precondition for being quoted on it. SiftServe’s own uplift figure is a hypothesis we test in pilots, not a result we claim.

Where to start

Pick your most-visited product page. Fetch it as text, cover the images, and list every buyer-relevant fact that disappeared. Write each one back in as a sentence or a table row with its units and conditions intact. Re-fetch, and confirm the facts are now in the text. That’s one page in an afternoon, and it’s the same check we run before any page of our own ships.

If the page can’t change, the AI-readable version is where the text goes, and it still goes through your approval before anything is served.