Images in AI Search: What Microsoft's Retail Guide Asks For

In December 2025 Microsoft Advertising published a sixteen-page guide for retailers called From discovery to influence: A guide to AEO and GEO. It argues that competition is moving from being found to being chosen, and it lists what a retailer should do about it.

It is worth reading, and it is free: from-discovery-to-influence-a-guide-to-aeo-and-geo.pdf.

Two things to know before you do. It is a Microsoft Advertising marketing asset, authored by a Microsoft Shopping and Copilot product manager, and it closes by inviting you to contact your account representative. And despite how it is often summarised, it is about Copilot and Bing, not ChatGPT. Its own key-takeaways page says so: “data signals that influence Copilot and Bing ranking.”

The advice is still sound. It is also, once the acronyms come off, ordinary technical SEO.

What it actually says about images

The paper’s action section runs to three strategies. Images appear in exactly two bullets:

Write detailed alt text and ImageObject schema describing visuals (“green jacket with reinforced zipper and extended hood”).

Provide video transcripts.

That is the whole of it. For a document about being understood by machines that increasingly reason over pictures, two bullets is thin. But both are right, and the parenthetical is the most useful thing in the paper.

The example is the lesson

Compare the two ways a retailer might describe the same photograph:

<!-- What most catalogues ship -->
<img src="jacket-green-01.jpg" alt="Green jacket">

<!-- What Microsoft's example asks for -->
<img src="jacket-green-01.jpg"
     alt="Green jacket with reinforced zipper and extended hood">

The first tells a model the category. The second tells it three attributes it can match against a question. Someone asking an assistant for “a rain jacket with a hood that covers my face” cannot be matched to the first and can be matched to the second.

This is the same alt text advice that has been correct for twenty years, for a different reason. It used to be about screen readers and Google Images. Now it is also the text a language model reads when it cannot see your picture.

The four things the paper leaves out

The image section reads like it was written by people who work on feeds rather than on images. Four gaps matter.

1. ImageObject without dimensions is half a schema

The paper names ImageObject and stops. In practice the useful properties are the ones that let a consumer decide whether your image is worth fetching:

{
  "@type": "ImageObject",
  "contentUrl": "https://example.com/jacket-green-01.jpg",
  "width": 2400,
  "height": 1600,
  "caption": "Green jacket with reinforced zipper and extended hood",
  "representativeOfPage": true
}

width and height should be read from the file, not typed from memory. We found that out on this site: our guide template declared every cover image as 1200×630, and only 84 of our 123 cover files actually were. Twenty-two published guides reserved the wrong box, and the page shifted as the image arrived — the exact Cumulative Layout Shift failure we tell other people to avoid.

Measuring the file at build time fixed both the markup and the schema at once. If you assert a dimension, read it.

2. Nobody checks what a crawler receives

Modern image delivery negotiates format from the Accept header. Ask for the same URL as a browser that accepts AVIF, and you get AVIF. Ask as something that does not, and you get JPEG.

That is excellent for users and it means the format in your schema may not be the format anyone receives. If a CDN negotiates delivery for you, do not declare encodingFormat at all. An unverifiable claim in structured data is worse than a missing one.

It also means the sanity check is not “what is in my repository” but “what comes back over the wire”:

curl -s -H 'Accept: image/avif,image/webp,*/*' \
  -o /dev/null -w '%{content_type} %{size_download}\n' \
  'https://example.com/jacket-green-01.jpg?w=1200'

Run that as an assistant’s fetcher would, not as your browser does.

3. A file extension is not a format

While auditing this site we found 39 files named .webp that are PNGs. Nothing was broken for readers, because the CDN transcodes on delivery and users receive AVIF at around 40 KB. But every tool that trusts the extension — including our own — reported the wrong format.

If you are enriching data so machines can trust it, the file name is data too. Detect format from the bytes or the Content-Type header, never from the characters after the dot.

4. Decorative and meaningful images need different treatment

“Write detailed alt text” is not true of every image. An image that carries meaning needs a description. An image that is pure decoration should have alt="", so assistive technology skips it instead of reading a caption that adds nothing.

The paper does not draw this line, and a retailer following it literally will write alt text for background textures and spacer graphics. That makes pages noisier for screen reader users and adds nothing for a model. Decide per image whether the picture carries information the surrounding text does not.

What the paper gets right, and nobody does

Buried in the takeaways is the sentence that justifies the whole document:

Retailers already hold most of the data signals that influence Copilot and Bing ranking — they’re just not surfaced in product feeds.

That is the honest core. Almost nothing in the action list requires new information. You already know your product’s dimensions, materials, and use case. The work is exposing what you know, consistently, in the three places that must agree:

1. The product feed you send to a merchant centre.
2. The structured data on the page.
3. The visible page a human reads.

The paper’s sharpest instruction is about that last agreement, and it is the oldest rule in search:

Ensure rendered DOM contains the same facts consumers see — never serve different HTML to bots.

That is the prohibition on cloaking, restated for a new audience. It has been a rule since the 1990s. If your answer to AI search involves showing a crawler something a person does not see, you have reinvented the thing search engines have penalised the longest.

A checklist that fits on one screen

For images specifically, the paper’s advice plus what it omits:

Do	Why
Write alt text that names attributes, not categories	It is the text a model reads when it cannot see the picture
Give decorative images alt=""	Fewer false descriptions for screen readers and models
Emit ImageObject with contentUrl and a real caption	Tells a consumer what the picture shows
Read width and height from the file	Wrong dimensions cause layout shift and false schema
Omit encodingFormat when a CDN negotiates delivery	You cannot know what arrives
Detect format from bytes, not the extension	Extensions lie more often than you would think
Keep feed, schema, and visible page consistent	The one rule the paper is emphatic about
Transcribe video before marking it up	VideoObject on an untranscribed clip adds nothing

On the acronyms

The paper introduces Answer Engine Optimization and Generative Engine Optimization as distinct disciplines. They are not distinct, and the paper’s own action list is one list, not two.

Treat “AEO” and “GEO” as Microsoft’s framing in a Microsoft sales document rather than as established terms of art. What is actually being described is search optimisation for systems that answer instead of listing. The techniques are valid schema, consistent data, honest markup, and descriptions written for a reader who cannot see.

The vocabulary is new. The work is not.

Read the source

The paper is sixteen pages, about 2,200 words, and free with no gate. It takes fifteen minutes. If you find it summarised in a thread that sells you an optimisation package between the numbered points, read the paper instead.

From discovery to influence: A guide to AEO and GEO — Microsoft Advertising, December 2025.