Optimizations
The five enhancements AI Pages applies, exactly what each one adds to the HTML, and when to switch one off.
Five enhancements make up the AI Pages transformation. Each one attacks a different way raw HTML fails a language model: no signal about what kind of page this is, numbers buried in paragraphs, an FAQ with no question-and-answer structure, the substance sitting below three screens of marketing, and proper nouns with nothing tying them to what they are.
All five are on by default and you can toggle each one independently in the Features step of setup, or later from Settings on the AI Pages page. Defaults are the right starting point for almost everyone. Read this page when you want to know what a specific feature puts on the page before you decide to turn it off.
How the transformation runs
The order matters, because everything downstream depends on the first two steps.
Trakkr strips what a model will not read: analytics and tag-manager scripts, and embedded snippets for tools like Hotjar, Segment, Mixpanel, Intercom, Drift and HubSpot. Then it classifies the page against nineteen types (homepage, product, pricing, blog post, FAQ, how-to, comparison, case study, local business, legal and others) and reads any JSON-LD you already publish.
That classification builds a single prompt that asks for everything your enabled features need at once, so one model call produces the schema, the facts, the questions, the summary and the entities together. The response is validated, merged with your existing schema, and injected.
If the model call fails, nothing is served. The optimization fails closed and your original page is what the crawler gets.
Two places carry the output. Head additions are meta tags and JSON-LD. Body additions are visible semantic sections appended at the end of the body, all inside a single <div data-prism-content="true"> so the whole insertion is one identifiable block. Every value is HTML-escaped, and the body block is discarded entirely if it somehow contains a <script> or <iframe>.
The five optimizations
Structured data injection
Generates JSON-LD for the page type it detected and adds it to the <head> as a <script type="application/ld+json">. A product page gets Product with an Offer, a blog post gets Article, a category page gets CollectionPage with an ItemList, a shop or clinic gets LocalBusiness with an address and geo point.
It helps because a model reading Nike Air Zoom Pegasus 40 $130 in body copy has to infer which token is the product, which is the brand and which is the price, and it has to do that against every other number on the page. Schema removes the inference.
Leave it off only when you have hand-built unusual schema and want nothing added beside it.
Key facts extraction
Pulls out the concrete data points on the page (prices, percentages, dates, counts, ratings, measurements) and publishes them twice: as an ai-key-facts meta tag holding up to ten facts as JSON, and as a visible Key Facts list in the body. Facts shorter than ten characters are dropped.
It helps most with accuracy. Models that invent a price or a date usually do so because they could not locate the real one. A short, explicit list of the page's own numbers is harder to miss and harder to substitute.
Leave it off for purely qualitative pages, a manifesto or a values page, where a bulleted fact list would read as odd. Keep it on for anything data-dense.
Automated FAQ injection
Derives question and answer pairs from what the page actually says and adds them two ways: as FAQPage JSON-LD in the head, and as a Frequently Asked Questions section of <details> blocks in the body. If the generated schema graph already contains an FAQPage, the separate FAQ script is skipped so the page does not carry two.
How many questions depends on the page type. Most types aim for three to five. A pricing page aims for four to six because pricing questions convert. An actual FAQ page extracts every pair on the page and aims for at least five. Contact, event and review pages get two or three. Legal pages get none.
Pairs are only kept when the question is at least ten characters and the answer at least twenty, which quietly drops the thin ones.
FAQPage markup with this one off.AI summary block
Writes a short factual summary of the page with the promotional language taken out, and publishes it as an ai-summary meta tag and a visible Summary section. Summaries under twenty characters are dropped.
It helps when your substance sits below the fold of a model's attention. If the opening of your page is a hero, a proof bar and three calls to action, a crawler that reads only the first stretch of a document never reaches what you actually do. A clean summary gives it something to hold.
Leave it off when the page is already structured around its substance, such as a long article or a documentation page, where the summary is redundant.
Entity recognition
Identifies the named things on the page (people, companies, products, technologies, places) and publishes them as ai-entities-* meta tags, up to twenty names per type, and as a Related Entities definition list in the body, up to fifteen names per type.
It helps most on comparison and category content. If your page names ten companies, an explicit list of who is who makes it far less likely that a model blends one brand's attributes into another when it writes an answer.
There is rarely a reason to turn it off. The one edge case is content that is deliberately ambiguous for stylistic reasons.
Turning one off
Each toggle is independent, and switching one off does not affect the others. What drops is cumulative legibility, so change one thing at a time and for a reason: brand voice, a regulated page, or a clash with markup you already control.
Test before you commit. Test Your Pages on the Overview tab takes any URL on your domain and runs the analysis without touching the live cache. It reports the detected page type, a schema quality score, which schema types were applied, how much content was added, and the HTML snippet for each enhancement. Compare the result with the feature on, then with it off.
Feature changes are server-side and do not need a redeploy, but they only reach a given URL when its cache entry regenerates, which is within seven days.
Turning all five off leaves the transformation with nothing to add. AI Pages then serves the stripped page with only a prism-optimized marker in the head.
What it will not do
- Invent facts. The transformer restructures what is on the page. If a number is not there, it is not in the optimized version either.
- Translate. The page comes back in the language it was written in.
- Change your messaging. Tone, positioning and narrative order are untouched. The transformations are structural.
- Pull in outside sources. Nothing is added from anywhere but the page itself, including citations and links.
To change what the optimized page says, change the source page. AI Pages is a delivery layer, not a content engine.
Common questions
Does the FAQ generation make things up?
It writes new sentences, but only about content that is already on the page, and the prompt tells it to create questions the content actually answers. The real risk is tone rather than fact, which is why the length guards drop thin pairs and why regulated pages are worth reviewing in Test Your Pages before you leave the feature on.
Do I need this if my page already has schema markup?
Structured data injection reads your existing JSON-LD and merges rather than replaces, so it fills gaps instead of fighting you. If you would rather own the schema entirely, turn that one feature off and keep the other four.
Are the added sections visible to people?
No. The body sections only exist in the version served to matched AI crawlers. A person visiting the same URL gets your normal page, with no Summary, Key Facts, FAQ or Related Entities section on it.
Where do the additions appear in the HTML?
Meta tags and JSON-LD go in the <head>. The visible sections are appended at the end of the <body> inside one <div data-prism-content="true"> wrapper, which is also how you or an auditor can identify exactly what was added.
How long until a feature change shows up?
On the next cache regeneration for each URL, which is within seven days. There is no manual cache invalidation, so a change rolls out page by page as entries expire.