AI crawlers do not rank formats against each other. Google's AI features find material by relying on core Search ranking systems to retrieve relevant, up-to-date web pages from the Search index, and the model then works from what it retrieved. Text is what it reads and quotes. Images and video get in through the words attached to them, and structured data labels what things are.
There is no format score
No AI company publishes a weighting that says video beats text, or a table beats a list. Searches for one turn up guesses. What Google does publish is how its AI features gather material, and that is more useful.
Generative AI features on Google Search rely on retrieval-augmented generation (RAG), a technique that improves the quality, accuracy and freshness of AI responses by relying on core Search ranking systems to retrieve relevant, up-to-date web pages. On top of that, query fan-out generates concurrent, related queries to fetch additional relevant search results for the one question the person asked.
So the first question for any format is not "how much is it worth". It is "can retrieval find this and read it". A video with no transcript, on a page with no text, gives the system very little to match against a question. A plain paragraph that answers the question gives it something to retrieve and something to quote. If the retrieval step is new to you, start with what RAG is.
Who is crawling, and why it matters for formats
Before formats, it helps to know who is reading. Bots fall into broad groups.
Crawlers for search indexing include GoogleBot, BingBot, YandexBot, Google Images, and Applebot. Crawlers that collect pages for training models include Amazonbot, GoogleOther, GPTBot (OpenAI), and PetalBot (Huawei).
AI answer engines sit on top of search. Perplexity, for example, is described as a tool connected to search engine APIs that can provide specific sources for generated text. Google's own AI features retrieve from the Search index, as above.
This changes how you think about formats. A training crawler collects text in bulk for a future model. A search crawler indexes a page so it can be matched to a question later. The second is the one that decides whether you are cited today, and it works through the same index that ranks ordinary results.
One more caution: some AI crawlers do not respect robots.txt guidelines. Robots.txt is a request, so it is a poor tool for controlling which formats a bot reads. More on the rules you can set in AI crawler rules.
Text: what gets read and quoted
Every AI system can read text, and an AI answer is text. So the passage that answers the question is the unit that matters. Put the question in a heading, put the answer in the first sentence under it, and keep the section about that one question. That gives retrieval a clean match and gives the model a sentence it can lift without rewriting.
Structure in the HTML carries through. One practitioner roundup notes that AI Overviews pull tables, images, videos, lists, alt text, and HTML structure from content. Real headings, real lists and real tables are easier to pull from than the same content styled to look like them.
What the text says matters more than how it is laid out. Google is direct about commodity content: a page like "7 Tips for First-Time Homebuyers" is often based on common knowledge which could originate from anyone and typically adds little unique insight. A page that says something only you can say, from your own data or your own product, has a reason to be the one cited.
And do not try to cover every phrasing with its own page. According to Google, creating separate content for every possible variation of search queries primarily to manipulate rankings or generative AI responses violates its scaled content abuse spam policy.
A practical check: if a page's text only appears after scripts run in the browser, a crawler that does not run scripts sees an empty page. Server-rendered or static HTML removes that risk.
Images: read through their words
Google says its generative AI search features can bring in relevant images and video, which means more opportunities for websites to appear beyond web page links. So images can show up in AI answers. But deciding whether an image belongs with a question still runs mostly through the words around it.
Those words have a name. Media metadata is descriptive information attached to images and videos, including captions, alt text, file names, transcriptions, subtitles, and geotags. Alt text is descriptive text placed in the img tag explaining the content of the image, and serves as an accessibility feature for visually impaired users.
What to do with that:
- Write alt text that says what the image shows, as you would describe it to someone who cannot see it. Not a list of keywords.
- Name the file for what it is. "ai-crawler-visits-by-page.png" says more than "screenshot-final.png".
- Caption screenshots with the point they prove, in a full sentence.
- If a chart carries a key number, put that number in the text next to it, with its source. A number inside an image cannot be quoted.
Video: transcripts and captions
How AI systems process video internally, whether they sample frames, read captions or both, is not documented by the engines. Anyone who tells you the exact method is guessing.
What is in your control is the text. Transcriptions and subtitles are part of a video's media metadata, and a transcript on the page turns a video into something retrieval can match and a model can quote. Google's AI features can bring in relevant video, so the video itself can appear too.
For each video worth being found:
- Publish the transcript on the same page, as text, not only as a caption file.
- Write a short summary above the video that answers the question the video answers.
- Add chapter titles that read like the questions people ask.
Structured data: telling the system what things are
In general, structured data means standardized and formatted data, including entries in spreadsheets, tables, databases, or fields of a content management system. On a web page, it usually means schema markup describing what the page is and what is on it.
For AI search, the claim is that structured data enhances entity recognition and retrieval for content with LLMs, making it more visible to GenAI searches. It tells the system that this is an article, this is a product, this is a question with this answer. The same roundup as above recommends Article, BlogPosting, HowTo, and FAQPage schema for AI Overviews.
Around the markup sits ordinary SEO metadata, and it still counts: page titles, meta descriptions, URL slugs, and canonicals, with meta robots, Open Graph tags, and hreflang as additional fields.
Treat structured data as labeling. It describes the page. It does not replace the text, and markup that claims something the visible page does not say is a mistake, not a shortcut.
Do ChatGPT, Claude, Gemini and Perplexity weigh formats differently?
None of them publishes how it weighs formats, so any ranking of formats by engine is a guess.
What is known is that they retrieve from different places. Google's AI features use the Search index and its ranking systems. Perplexity connects to search engine APIs. Because retrieval differs, the same page can be cited by one engine and ignored by another, whatever its format.
The way to find out for your pages is to test. Ask each engine the same buyer questions, several times, and record which pages and which formats get cited. That repeated check is a visibility run. If the engine you care about keeps citing a competitor's comparison table and not your video, that tells you more than any general rule.
A checklist by format
| Format | What the AI system reads | What to do |
|---|---|---|
| Text | The words, headings and lists in the HTML | One question per section, answer in the first sentence |
| Images | Alt text, caption, file name, nearby text | Describe what it shows; put key numbers in text |
| Video | Transcript, subtitles, the text around it | Publish the transcript on the page |
| Tables | Rows and columns as HTML | Use a real table element, not an image of one |
| Structured data | Schema properties | Mark up what the visible page says, nothing more |
How to check what AI crawlers fetch
You cannot see what a model does with your page, but you can see what its crawler fetched. Log file analysis can show which crawlers access a site, how frequently they visit, and what pages they prioritize. Botify suggests watching crawl frequency, crawl depth (how many layers of a site are being crawled), and crawl budget (the number of pages crawled in a given time frame).
Look at it by format. Do AI bots fetch your image files, or only your HTML? Do they reach the pages that hold your transcripts? A page no bot visits cannot be retrieved.
Links help bots get there. In one case study, pages that were not getting any AI crawl activity started appearing in AI Overviews after adding contextual internal links from high-authority pages.
Bots do not run analytics scripts, so this needs server logs or a server-side counter. More in how to know if AI crawlers are hitting your site. If you build with Claude Code or Codex, Shipfound counts GPTBot, PerplexityBot and ClaudeBot visits per page and checks in repeated runs whether the AI engines name your product.
Frequently asked questions
How do I optimize for AI crawlers?
Allow the search bots in, put every answer in text on the page, describe images and video in words, add structured data that matches the visible page, and link to new pages from pages that already get visits. Then check your logs to confirm the bots arrive.
How do I rank in AI search?
Start with ordinary search. Google's AI features use core Search ranking systems to retrieve the pages they draw on, so a page that cannot rank is unlikely to be cited. Then cover the related questions a buyer would ask, because one question becomes many searches. More in AI Overviews ranking factors.
Do AI crawlers read images?
They fetch them, and Google's AI features can bring in relevant images. Whether an image fits a question is judged largely from the text attached to it: alt text, caption, file name and the words nearby.
Do AI crawlers watch videos?
The engines do not document how they process video. A transcript published on the page is the part you know they can read.