If you add a fact to an image filename, alt text, or caption, but don’t mention it anywhere in the visible body text, can AI search engines pick it up? We tested it across five AI search platforms using six page variations, and the answer was clear: image metadata alone is not enough.
This experiment is part of OtterlyAI’s ongoing GEO research series, where we test specific content signals to understand what AI search platforms can and can’t read.
Key Findings (TL;DR)
Image metadata alone does not reliably inform AI search. No single image signal (filename, alt text, caption, or text rendered inside an image) produced correct answers on any platform.
55% Hallucination Rate: When AI search can’t find the information, most AI platforms will make up an answer instead of saying “I don’t know.” which is a bigger concern for brands that have incomplete information about their products/services.
AI platforms didn’t seem to read text baked into images. None of the five platforms extracted the fact from Page V4, where the answer was rendered as a text overlay, questioning if infographics, screenshot tables, and chart labels are invisible to AI search?
Only partial success came from combining filename + alt text. Only ChatGPT, which returned the correct answer in 75% of attempts for that specific variation.
Copilot Most Honest Platform. It cited the right URLs consistently but never returned the correct answer, and never hallucinated when it couldn’t extract the information.
The negative control page (no signal at all) still triggered hallucinated answers on three out of four platforms that could access the pages, confirming that these platforms will fabricate information when it isn’t present.
Why This Experiment Matters for GEO
Image optimization is a standard part of SEO. Descriptive filenames, accurate alt text, and captions are common best practices. But in the context of GEO, the question is different: can AI search engines extract factual information from these signals when it doesn’t appear in the page’s body text?
This matters for any brand that relies on images to convey information, whether it’s product specifications in image captions, team details in photo descriptions, or data embedded in infographics. If AI platforms can’t read these signals, that information is invisible in AI-generated answers.
Scope of Study
This experiment was run on a single test domain with existing content. The findings are specific to the platforms and prompts tested during the one-week experiment window.
We tested five AI search platforms: ChatGPT, Google AI Mode, Perplexity, Gemini, and Microsoft Copilot. Each page variation was tested multiple times, for a total of 120 individual test runs across all platforms and variations.
We needed a fact that doesn’t exist anywhere on the internet, so we could trace exactly where an AI platform got its answer. We used a fictional team statistic about OtterlyAI: the number of vegetarians on the team.
Each of the six test pages used a different number as the “correct” answer. If an AI platform returned a specific number, we could identify which page (and which signal) it extracted the information from. If it returned a number that didn’t match any page, we knew it was hallucinating. To help the AI platforms find the information, these pages were added to the footer of the website for easy reference.
Testing for OCR (Optical Character Recognition)
One page (variation #4) was designed specifically to test whether AI search platforms can use OCR, the technology that lets a computer read text inside an image. Instead of placing the vegetarian fact in a filename, alt text, or caption, we baked the number directly into the image itself as a text overlay on the photo. The filename was generic, no alt text was added, and the body content never mentioned the fact. If a platform returned the correct number, it would prove the AI could “read” pixels the same way it reads HTML. If not, it would confirm that text rendered inside an image is invisible to AI search retrieval.
The Six Page Variations
All six pages used identical body text: a short article about OtterlyAI’s 17-person team, the founding story, milestones, and company background. The body text never mentioned vegetarians. The only place the vegetarian fact existed was in the image signal being tested.
Page variation
Where the Fact Was Hidden
1. Image filename only
– The fact was embedded in the image filename (e.g., 1-vegetarian-in-otterlyai-team.jpg). – No alt text, no caption, no body mention.
2. Alt text only
– The fact was in the image alt attribute. (e.g., 2 vegetarians are in the OtterlyAI team). – Generic filename. – No caption, no body mention.
3. Caption (figcaption) only
– The fact appeared as a visible caption below the image using a <figcaption> element. (e.g., 3 vegetarians are in the OtterlyAI team). – Generic filename. – No alt text, no body mention.
4. The OCR Image Test
– The fact was baked into the image pixels (text overlay on the photo). Tests OCR/vision capability. – Generic filename -No alt text, no caption, no body mention.
5. Nothing
– No signal at all. As a negative control test. The correct behavior here is “I don’t know.” – Generic filename, – No alt text, no caption, no body mention.
6. Filename + alt text combined
– Both the filename + alt text contained the fact (e.g., 6 vegetarians are in the OtterlyAI team). – No caption, no body mention.
Monitoring Setup
We ran each prompt multiple times across all five platforms over the course of one week. Each run asked the AI platform about the vegetarian count on OtterlyAI’s team.
2 different brand reports were used in OtterlyAI:
One report had prompts without specific pages:
The other pointing it to the specific test pages mentioned above:
For each test run, we recorded three things:
The answer given (the number returned, if any)
Whether the answer was correct (matched the number on the tested page)
Which page(s) the platform cited (to check if it even looked at the right page)
Results
Overall Accuracy: 2.5%
Out of 120 total test runs,only 3 returned the correct answer. All three came from a single combination: ✅ Filename + alt text combined had the highest correct answers but only on ChatGPT.
Every other combination of page and platform either hallucinated, gave a wrong answer, or couldn’t access the page (due to the page not yet being indexed in its memory)
Results Matrix
ChatGPT
Google AI Mode
Perplexity
Gemini
Copilot
Filename only
⚠️ Hallucinated
⚠️ Hallucinated
❌ Wrong
❌ Wrong
❌ Wrong
Alt text only
⚠️ Hallucinated
⚠️ Hallucinated
❌ Wrong
⚠️ Hallucinated
❌ Wrong
Figcaption only
⚠️ Hallucinated
⚠️ Hallucinated
❌ Wrong
⚠️ Hallucinated
❌ Wrong
OCR image
⚠️ Hallucinated
⚠️ Hallucinated
❌ Wrong
⚠️ Hallucinated
❌ Wrong
No signal / control
⚠️ Hallucinated
⚠️ Hallucinated
❌ Wrong
⚠️ Hallucinated
❌ Wrong
Filename + alt text
✅ Partially correct (75%)
⚠️ Hallucinated
❌ Wrong
⚠️ Hallucinated
❌ Wrong
Legend:
⚠️ = Hallucinated (gave a confident wrong answer)
❌ = Wrong or no access (did not hallucinate /fabricate)
✅ = (Partially) Correct
OCR Test: No Platform Could Read Text Inside the Image
Variation #4 produced zero correct answers across all five platforms. ChatGPT, Google AI Mode, Gemini, and Copilot treated the page the same way they treated the negative control: they either hallucinated a number or returned nothing at all. Perplexity, as with every other variation, could not access the page.
This confirms that AI search platforms (in non-deep thinking mode) are not using OCR to extract factual information from images during retrieval, even when the text is clearly rendered, high-contrast, and readable to a human.
Example used for this experiment. (A minimal design was used to avoid confusing the AI):
For brands, this matters whenever key information is delivered visually: pricing tables saved as images, product specs inside infographics, team details written across a photo, or data labels on a chart screenshot. From AI search’s perspective, that content does not exist.
ChatGPT was the only platform to return correct answers, and only for the version where the image had both the correct answer in the filename + alt text combined, where it got the right number in 75% of attempts. For every other variation, ChatGPT hallucinated. It returned confident numbers that didn’t match any test page, sometimes pulling information from completely unrelated pages.
Shows 75% a correct answer if both image filename + alt text are present.
ChatGPT also frequently looks at an average of 21 different pages on and off the website (even when instructed that it should look at 1 specific page). When asked about one variation, it would sometimes cite a different variation’s URL, mixing up sources across the test set.
Results: Hallucination rate: 83% of all ChatGPT test runs.
2. Google AI Mode: 100% Hallucination
Google AI Mode hallucinated on every single test run across all six page variations. It never returned a correct answer. Like ChatGPT, it frequently cited the wrong test pages and mixed up sources. Even for Page V6 (the combined signal), AI Mode returned wrong numbers in every attempt.
Results: Hallucination rate: 100% of all AI Mode test runs.
3. Perplexity: Zero Access
Perplexity could not access any of the test pages. It returned no answers across all variations. This is consistent with findings from other OtterlyAI experiments, where Perplexity’s access to smaller or newer domains can be limited.
This might be caused by the way Perplexity updates and indexes new pages and websites to its memory.
Results: Access rate: 0%.
4. Gemini: 100% Hallucination
Gemini hallucinated on nearly every test run. It returned confident but incorrect numbers for all page variations, including the negative control page (V5), where no information existed at all. (We don’t have a Nadine in our team at all)
Gemini cited the correct pages less frequently than other platforms, often returning answers with no source citations.
Results: Hallucination rate: 100% of all Gemini test runs.
5. Copilot: Didn’t have the correct answer but didn’t hallucinate
Copilot’s behavior was the most interesting. It consistently found and cited the correct test pages, especially for the filename (V1) and alt text (V2) variations. But it never extracted the vegetarian count from any image signal. It returned “unknown” or zero across the board.
Copilot also never hallucinated. When it couldn’t find the information, it didn’t make up an answer. This is a notable difference from ChatGPT, AI Mode, and Gemini, which all fabricated numbers with confidence.
Results: Hallucination rate: 0%. Correct answer rate: also 0%.
What This Means for GEO
Image Metadata Alone Is Not a Reliable Content Signal for AI Search
The core finding is simple. If a fact exists only in image metadata (filename, alt text, caption, or text rendered inside an image), AI search platforms will not reliably extract it. Out of 120 test runs across five platforms, the correct answer appeared only 2.5% of the time.
For GEO practitioners, this means that any information you want AI platforms to surface needs to be in the visible, crawlable body text of your page. Image signals are not a substitute or a method to hide hidden information/instructions to the AI Search platform.
Filename + Alt Text Is the Strongest Image Signal (But Still Weak)
The only combination that showed any success was filename plus alt text together, and only on ChatGPT. This suggests that when multiple image attributes reinforce the same information, there’s a marginally higher chance of extraction. But “75% accuracy on one platform” is not a strategy you can rely on.
Captions Don’t Perform Better Than Hidden Metadata
A surprising finding: the visible <figcaption> caption (V3) performed no better than the hidden filename (V1) or alt text (V2). Despite being visible text on the rendered page, the caption did not produce correct answers on any platform.
One possible explanation: AI crawlers may not treat <figcaption> text the same way they treat paragraph or heading text. The semantic association with the image (via the <figure> element) may not carry the same weight as standalone body content.
AI Platforms Cannot Read Text Inside Images
Page V4 tested whether AI platforms could use OCR or vision capabilities to read text rendered inside an image. None of them could. This confirms that text baked into images (infographics, screenshots with overlaid text, charts with labels) is invisible to AI search retrieval.
If your content includes information in image form (pricing tables as images, product specs in infographics, data in chart screenshots), that information does not exist for AI search purposes unless it’s also written out in the body text.
The Hallucination Problem Is Significant
Over 55% of all test runs produced hallucinated answers. Platforms returned specific numbers with confidence, even on the negative control page where no information existed at all. Three out of four platforms that could access the pages hallucinated on the control page.
This reinforces a pattern from our previous hidden text experiment: when AI platforms can’t find information, they often fabricate it rather than acknowledging the gap. For brands, this means AI could be generating incorrect facts about your products, your team, or your positioning, and presenting them as truth. In this previous experiment’s example, the Google AI Mode cited an non-existent poem:
Wrong Page Citations Are Common
Beyond hallucinating answers, platforms frequently cited the wrong source pages. ChatGPT and Google AI Mode both had instances where they cited a different test page than the one specifically being instructed to look at. This means even when a platform cites a URL in its answer, the information may not have come from that page.
For brands monitoring AI search, this makes citation tracking more complex. A citation to your website/page doesn’t guarantee the AI extracted information from it, and information attributed to your page may have come from somewhere else entirely.
Practical Takeaways
1. Put critical information in body text, not in images. If you want AI search to surface a fact, write it in a paragraph or heading. Don’t rely on image filenames, alt text, captions, or text rendered inside images.
2. Don’t stop optimizing images for SEO. Descriptive filenames and accurate alt text still matter for traditional search, accessibility, and user experience. They don’t hurt your GEO. They are simply not sufficient as the only source of a fact.
3. Audit pages where key information lives only in images. Product pages with specs in image format, team pages with details only in photo captions, pricing pages using screenshot-based tables. These are all GEO blind spots. Add the same information as HTML text.
4. Monitor for hallucinated information about your brand. ❗If AI platforms are willing to fabricate specific numbers about a team’s dietary preferences, they’ll fabricate other details too. Use OtterlyAI’s Brand Reports and Brand Sentiment Analysis to track what AI platforms say about you and catch misrepresentations early.
5. If you combine image signals, filename + alt text is your best bet. It’s the only combination that showed any success, and only on one platform. Use it as a reinforcement of body text, not as a replacement.
Pro Tip: Explore ourGEO experimentation sheet to track live updates on the latest AI Search studies.
Closing Thoughts
This experiment tested a simple question: can AI search platforms extract factual information from image signals when it’s not written in the body text? The answer, across 120 test runs and five platforms, is no.
Image filenames, alt text, captions, and text rendered inside images are not reliable content signals for AI search. The only partial success came from combining filename and alt text on a single platform (ChatGPT), and even that worked only 75% of the time.
❗The bigger concern is what happens when AI platforms can’t find any answer at all for a specific user-query. And how certain are you that your website contains all the information needed to provide an answer to AI platforms? In this simple experiment, over half the time when AI didn’t find an answer to the question, they made one up. They returned specific, confident numbers that matched no source. They cited pages they didn’t extract information from. And they did this even on a control page that contained no information at all.
For GEO, there are 2 clear takeaways:
Image meta data alone doesn’t cut it. If a fact isn’t in your body content as crawlable HTML, it doesn’t exist for AI search. Image optimization is good practice, but it’s a supplement to text, not a substitute for it.
This experiment reveals an even bigger problem; the results should worry every brand. If your website is missing exact match information, there is a high chance that AI search platforms will fabricate things. (In this experiment that was 55% of the time) → FAQ-pages should be the bare minimum for brands that want to ensure what AI says is correct.
Have You Tested Image Signals on Your Own Site?
These findings are based on a single test domain with existing content. Image signal behavior may differ on established, high-authority domains or on pages with hundreds of images and robust structured data. If you’ve run your own tests, we’d like to hear about it.
Share your experiment results with the OtterlyAI team at rick.tousseyn@otterly.ai. We may feature your findings in a future update to this research.
