How to Optimize for AI Image and Multimodal Search in ChatGPT and Gemini

To win AI image and multimodal search in ChatGPT and Gemini, pair every image with descriptive alt text, a keyword rich filename, a caption, and ImageObject or Product schema, submit a complete image sitemap, and keep pages crawlable by AI bots. These signals let vision models retrieve, understand, and correctly attribute your images instead of a competitor's.

*By Josh Bernstein | August 16, 2026*

ChatGPT now answers more than 900 million people every week, according to OpenAI's February 2026 disclosure, and a growing share of those conversations start or end with a photo rather than a typed query. OpenAI's shopping expansion lets someone upload a picture of a product and get matched listings from connected retailers. Google is moving the same direction from the other side: Google has said Lens alone handles close to 20 billion visual searches a month, and its February 2026 Circle to Search update, powered by Gemini 3, now lets a single photo of an entire scene trigger results for every object in it at once, not just one circled item.

For brands that spent the last two years optimizing text for GEO, image and multimodal search is the uncovered flank. Clicks are already harder to earn everywhere: AI Overviews cut click-through rate on the top organic result by 58% as of December 2025, up from a 34.5% drop in April 2025, according to Ahrefs research. When ChatGPT, Gemini, or Google Lens pulls in an image instead of sending someone to your page, that image, and everything wrapped around it, is doing the job a headline and meta description used to do. This guide covers how multimodal retrieval actually works, and the specific alt text, filename, sitemap, and structured data moves that make your images retrievable, attributable, and, where the platform allows it, clickable.

Why Multimodal Search Is Already a Visibility Problem You Can Measure

Image and multimodal search is not a future trend to plan for later. It is a present-tense channel that is already shifting spend and attention in the GEO market. Sitecore acquired the GEO monitoring startup Scrunch for roughly $225 million in June 2026, and Profound raised a $96 million Series C at a $1 billion valuation in February 2026, both explicitly to help brands track and influence how they appear across AI platforms, images included. That kind of capital does not chase a niche.

The underlying behavior shift is straightforward: search is becoming less about typing keywords and more about pointing a camera or dropping a screenshot into a chat window. Google's own product roadmap makes the point for you. The redesigned Google Images experience, which began rolling out in July 2026, replaced the classic results grid with a Pinterest-style discovery feed built for continuous, AI-assisted browsing rather than one-off lookups. Multisearch lets someone snap a photo of an item and add text on top of it, like photographing a dining table and typing "coffee table" to find a matching piece. None of that behavior existed in mainstream search five years ago, and none of it rewards a site that treats images as decoration.

The practical consequence for marketers is that image optimization has quietly become a citation and attribution problem, not just a page-speed problem. A product photo with a generic filename and empty alt attribute is functionally invisible to the systems now doing the discovery work on your buyers' behalf. Fixing that is not exotic. It is a return to fundamentals that most sites let slide once classic Google Image Search stopped feeling like a priority channel, combined with a handful of newer structured data and sitemap requirements that AI crawlers actually depend on.

How ChatGPT Vision, Gemini, and Google Lens Actually "See" an Image

It helps to separate two very different things that all get lumped under "AI can see images now."

The first is a user uploading their own photo directly into ChatGPT, Gemini, or Google Lens. In that case, a vision-language model genuinely processes the pixels: it detects objects, reads text in the frame, estimates materials and colors, and reasons about how elements relate to each other. Gemini 3's contribution to Circle to Search is a good example, letting the model recognize every garment in an outfit photo simultaneously instead of one item at a time. ChatGPT's shopping feature works the same way on the input side: upload a photo of a product from a screenshot or a real-world item, and the model identifies what it is before searching for matches.

The second, and far more relevant one for your website, is a model needing to find or reference an image that already exists somewhere on the open web in response to a query. Here, the model is not scanning the live internet pixel by pixel in real time. It is working from an index, and what gets indexed, ranked, and eventually surfaced depends overwhelmingly on the text layer wrapped around that image: the alt attribute, the caption, the filename, the surrounding paragraph, and any structured data on the page. Pixel understanding alone cannot tell a retrieval system which specific product this is, what page it lives on, whether the image is current, or whether it is even licensed to display. That context has to come from machine-readable text, which is exactly why classic image SEO fundamentals matter more, not less, in a multimodal search era.

*Six text-layer signals get merged with the image itself into one searchable representation, which is what ChatGPT, Gemini, and Google Lens actually match against a user's query.*

This is also why Google's AI Mode and Gemini integration matters so much for image queries specifically. AI Mode's multisearch takes a photo plus a text refinement and reasons about the whole scene, including how objects relate to one another, their materials, colors, shapes, and arrangement. That reasoning still needs an anchor on your side of the exchange: a product page with clean, specific, machine-readable descriptions of exactly those attributes.

Alt Text, Captions, and Filenames: The Text Layer That Makes Images Machine-Readable

Alt text is the single highest-leverage image optimization move because it is the one signal every major system, from classic Googlebot to GPTBot to Gemini's crawler, reliably reads and attaches directly to the image. Google's own Search Central documentation still frames the test simply: describe what information a person would miss if they could not see the image, then place the natural keyword where it genuinely fits, rather than stuffing it in.

Good alt text for AI retrieval purposes does more than describe what a human sees. It should also carry the specific, disambiguating details a model needs to match the image to a query: the product name, distinguishing attributes like material, color, or size, and the context in which it is being shown. "Blue running shoe" tells a model almost nothing useful. "Men's Cascade Trail running shoe in cobalt blue with a reinforced toe cap, side view" gives a retrieval system enough to confidently match it against a specific buyer question.

Captions deserve more attention than most sites give them. Unlike alt text, which is invisible metadata, a caption is visible running copy sitting directly next to the image, which means it carries strong contextual weight for both classic search and AI extraction. A caption that states a fact, a price, a use case, or a comparison point next to the image gives a model text it can lift almost verbatim into an answer, similar to how the Princeton and Georgia Tech GEO study found that adding concrete statistics lifted AI citation performance by as much as 41%, direct quotations by up to 28%, and citing authoritative sources by as much as 115% for lower-ranked pages.

Filenames are the smallest of the three signals but still worth fixing, because it costs nothing and Google explicitly confirms it helps. Use short, descriptive, hyphenated filenames that name the actual subject, the same logic Google gives in its own example of preferring my-new-black-kitten.jpg over IMG00023.JPG.

Apply these rules consistently across a site:

• Write one clear, complete sentence of alt text per image; never leave it blank on a content or product image, and use an empty alt="" only for genuinely decorative images.
• Lead with the specific noun (product name, place, person, concept) before adding descriptive modifiers.
• Avoid keyword stuffing or repeating the same phrase across every image on a page; each alt attribute should describe that specific image.
• Add a visible caption wherever an image sits inside content that AI engines are likely to extract from, not just on hero images.
• Rename image files before upload using lowercase words separated by hyphens, matching the primary subject of the image.
• Keep alt text and captions in sync with the actual image; mismatched or outdated descriptions erode trust signals over time as pages get re-crawled.

Image Sitemaps and Technical Delivery: If AI Crawlers Can't Fetch It, None of This Matters

None of the metadata work above matters if crawlers cannot reliably discover and fetch your images in the first place. An image sitemap gives search and AI crawlers a direct, structured map to every image worth indexing, separate from whatever a crawler happens to notice while rendering a page.

The technical requirements are specific and easy to get wrong. A sitemap needs the standard sitemap namespace plus the dedicated xmlns:image namespace; without it, image tags inside the sitemap are simply ignored. Within that structure, you can list up to 1,000 <image:image> entries per <url> block, which matters for product pages with multiple angles or content pages with large galleries. Image URLs can also point to a different domain than the page itself, which means sites serving images from a CDN do not need to fight their sitemap tooling to stay compliant.

Beyond the sitemap file itself, a handful of delivery details determine whether an image is even reachable:

• Confirm robots.txt is not blocking image crawlers, including Googlebot-Image and the crawler user agents used by AI platforms.
• Avoid loading key images purely through client-side JavaScript with no fallback; use native loading="lazy" rather than logic that only renders on scroll or interaction, since a crawler that does not fully execute your JS may never see the image tag.
• Serve images in widely supported formats (JPEG, PNG, WebP) and pair them with srcset or <picture> for responsive delivery, since page speed and mobile usability still influence how deeply a page gets crawled.
• Keep the image sitemap dynamic and auto-updating as content changes rather than manually maintained, so new product photos or article images do not sit unindexed for weeks.
• Resubmit and monitor the sitemap in Google Search Console, and check the Image Search performance report there periodically to confirm indexing is actually happening, not just assumed.

This is really a subset of a broader discipline. If your site has crawl budget problems, inconsistent rendering, or blocked resources, image visibility will be one of the first casualties, which is why technical GEO fundamentals around crawlability and clean HTML structure are worth auditing before investing heavily in image content itself.

Structured Data for Images and Products: ImageObject, Product, and What Still Matters

Structured data is where a lot of teams either overthink or under-invest. The distinction that matters is simple: ImageObject schema describes the image itself (its subject, license, creator, and caption as data), while Product schema describes what is being sold, with the image as one property among price, availability, and identifiers. Use ImageObject when an image is genuinely the point of the page, like a photography portfolio or an editorial visual. Use Product schema when the page exists to sell something, which covers the overwhelming majority of ecommerce use cases. Both can, and often should, appear on the same page.

For ecommerce specifically, structured data on the page and a feed submitted through Google Merchant Center are complementary, not redundant. Google has stated directly that providing both maximizes eligibility for shopping experiences and helps Google verify that on-page data matches the feed. This same underlying product data is what powers connections into newer surfaces: ChatGPT's shopping experience runs on the Agentic Commerce Protocol, which lets merchants, Shopify sellers among them, feed structured catalog data directly into ChatGPT so a photo upload can resolve into an actual matched listing with price and availability rather than a vague description.

It is worth being precise about one thing that trips people up: schema's payoff has shifted, not disappeared. HowTo rich results were removed from Google's visual search results back in 2023, and FAQ rich results were fully removed from Google Search as of May 7, 2026. Neither of those removals means the underlying markup stopped being useful. Both FAQPage and HowTo remain valid, fully supported schema.org types, and they continue to help AI systems parse a page's structure cleanly, matching a step to its accompanying image, or a question to its answer, which is exactly the kind of clean parsing that improves whether a model extracts your content (and the images inside it) correctly. Keep FAQPage and HowTo alongside Article or BlogPosting and ImageObject or Product markup. You should not expect a visual rich result from it anymore, but you should still expect cleaner AI extraction. For a fuller walkthrough of which schema types earn AI citations beyond images specifically, see how schema markup wins AI citations.

From Retrieval to Attribution: What Actually Gets Cited, Linked, or Silently Reused

Getting an image retrieved and getting it attributed back to you are two different outcomes, and most brands only ever achieve the first. An AI answer can absolutely use your product photo, your diagram, or your team headshot without a visible link back to your site, especially when the system is synthesizing a new visual layout rather than displaying a sourced image directly.

The three major surfaces behave differently enough that a single strategy will not cover all of them equally well.

Google Lens and Circle to Search remain the most attribution-friendly of the three because the underlying interaction is still fundamentally a results page, just triggered by a photo instead of typed text. ChatGPT's shopping surface sits in the middle: attribution depends heavily on whether a merchant's catalog is actually connected through the Agentic Commerce Protocol, which is a strong argument for optimizing product pages for ChatGPT Shopping and Gemini rather than assuming good alt text alone will get you featured. Gemini's AI Mode is the least predictable, since its multimodal reasoning increasingly favors generating a new visual answer over surfacing an existing image untouched.

It is also worth remembering that AI systems do not only pull images from brand-owned domains. Community platforms feed enormous volumes of user-uploaded product photos, comparisons, and real-world context into these systems, and Reddit in particular is already established as the single most-cited source across major AI engines for text answers. The same dynamic increasingly applies to images: a well-captioned photo in a Reddit thread or a detailed comparison post can end up shaping what a model shows or describes about your category, whether or not you uploaded it yourself. Treating owned-site image SEO as the entire strategy misses where a meaningful share of multimodal context actually originates. The same logic applies to comparison shopping behavior generally, which is why it is worth reviewing how Perplexity handles shopping and product discovery alongside your ChatGPT and Gemini work rather than treating each platform in isolation.

Building and Measuring an Image-Driven AI Visibility Program

Turning all of the above into an actual program is mostly a sequencing problem. Fix the highest-leverage, lowest-effort items first, then layer in the technical and structural work that compounds over time.

1. Audit your current image inventory on your highest-traffic and highest-revenue pages first. Pull a sample and check for missing alt text, generic filenames, and pages with no visible captions at all.
2. Rewrite alt text and captions on that priority set using specific, disambiguating language rather than generic descriptions, following the rules above.
3. Rename and re-upload critical image files with descriptive, hyphenated names, prioritizing product and hero images over minor decorative assets.
4. Build or repair your image sitemap, confirming the xmlns:image namespace is present and that robots.txt is not blocking image or AI crawlers.
5. Add ImageObject and Product schema where relevant, and connect an actual product feed (Merchant Center, or the Agentic Commerce Protocol for ChatGPT) if you sell anything online.
6. Test manually across platforms by running your own product and content photos through ChatGPT, Gemini, and Google Lens with realistic buyer questions, and note whether your brand appears, gets described accurately, or gets attributed at all.
7. Monitor on a recurring cadence, not as a one-time project, since model behavior and platform features are both changing quickly enough that a quarterly check is already too slow.

That last step is where most teams underinvest. Google Search Console's Image Search performance report shows classic image indexing and click data, but it will not tell you whether ChatGPT described your product correctly, showed a competitor's photo instead, or omitted your brand from a Lens-triggered shopping result entirely. Prompt testing by hand catches obvious problems but does not scale and is hard to prove to leadership as a trend. A structured monitoring cadence, tracked the same way you would track rankings or referral traffic, is what turns image and multimodal visibility from a guess into a number you can report on. If you are still reporting on GEO with ad hoc screenshots, it is worth reading how to build a reporting stack and KPI set for AI visibility that treats image and multimodal citations as a tracked line item alongside text citations, not an afterthought.

Frequently Asked Questions

Does alt text actually matter for ChatGPT and Gemini, or only for classic Google Image Search?

It matters for both, and arguably matters more now, because alt text is one of the few reliably indexed, machine-readable signals a retrieval system has to match an image to a query without re-analyzing raw pixels at scale. Google's own documentation still treats alt text as core image SEO guidance, and that same text layer is what AI crawlers rely on when deciding whether an image is relevant to a user's question.

Is HowTo or FAQ schema still worth using if the rich results are gone from Google?

Yes. The visible rich result in Google's search results page for both HowTo and FAQPage schema has been removed, but both remain valid schema.org types, and the underlying markup still helps AI systems parse your page structure cleanly, including matching steps to their images. Keep the markup for extraction quality even though it will not produce a visual SERP feature anymore.

What is the difference between ImageObject and Product structured data?

ImageObject schema describes the image itself, including its subject, caption, creator, and license, and is best used when the image is the actual subject of the page. Product schema describes an item for sale, with the image as one property alongside price and availability, and is the right choice for the vast majority of ecommerce pages. Both can appear on the same page without conflict.

How does ChatGPT's shopping feature actually find and display product images?

ChatGPT's shopping experience runs on the Agentic Commerce Protocol, which lets merchants, including Shopify sellers, feed structured product catalog data directly into ChatGPT. When a user uploads a photo or describes a product, ChatGPT identifies it and matches it against connected retailer catalogs, then displays images, prices, and specs in a visual, side-by-side format.

Do I need a separate image sitemap if my images are already in my regular XML sitemap?

You need the dedicated xmlns:image namespace and <image:image> tags either within your existing URL entries or in a standalone image sitemap; a regular sitemap without that namespace will have any image tags ignored entirely. It does not need to be a separate file, but the namespace and tag structure are non-negotiable for images to be properly indexed.

Can AI engines use my images without linking back to my site?

Yes, and this happens often. Being retrieved and being attributed are two different outcomes. Google Lens and Circle to Search tend to link back to source or listing pages because the interaction functions like a results page, while ChatGPT and Gemini can synthesize new visual layouts or describe an image without a direct citation, especially when no structured product feed connects your catalog to the platform.

How do I know if my images are actually showing up in AI answers?

Combine manual testing, running your own product and content photos through ChatGPT, Gemini, and Google Lens with realistic buyer prompts, with Google Search Console's Image Search performance report for classic indexing signals, and a recurring monitoring cadence rather than a one-time check, since platform behavior shifts frequently enough that quarterly spot checks miss real changes.

Should I prioritize alt text or structured data first if I only have time for one?

Start with alt text, captions, and filenames, since they are the fastest to fix, require no development work, and are read by every major crawler without exception. Structured data compounds that foundation and becomes especially important once you are selling products or publishing instructional content with step-by-step images.