What AISA reads on a page

When AISA reads a page, it keeps two things: the page's text, and its content, which is the same page sorted into facts, sections under their titles, and shopper reviews. The text feeds the writer of starter questions. The content is what the assistant reads when a shopper asks about the page, through the Site page tools.

AISA reads the HTML your server sends. It does not run JavaScript. See How AISA fetches your pages.

What the content holdsCopy Anchor Link

The 100 facts and the 100 FAQ facts are counted apart, so a long FAQ never pushes the specifications out. A fact past its limit is left out.

Storefronts that publish the product in their page stateCopy Anchor Link

Some storefronts send the product's specification table and FAQ only as data inside the page, not as text and not in the structured data. SAP Composable Storefront (Spartacus) does this: the data sits in a <script id="ng-state" type="application/json"> block. When a page has one, AISA reads the specification features and the FAQ entries of the page's own product and keeps them as facts, after the facts of the structured data. A feature whose only value is NA is skipped, and an FAQ answer is kept as plain text.

On an accessory page (a bowl, a lid, a filter), the page shows only the first few rows of the appliances it fits, while the state holds the whole list. AISA keeps that list as one or more sections titled Compatible products, after the page's own sections. Each entry is the product reference followed by its name, separated by ; . An entry with no reference is skipped, an entry with no name keeps the reference alone, and an exact repeat is kept once, so the count can be lower than the "Designed for N products" heading on the page. The list is placed in the budget before the page's own text, so a long page never pushes it out. Each section stays within the 4,000 character limit and ends on a whole entry.

A page read before this existed does not have the section. Read it again to get it.

AISA never keeps a price, a discount or a stock level from this data, and nothing about other products shown on the page. If the block is missing or not in the expected shape, nothing is added.

Shopify collections drawn in the browserCopy Anchor Link

Some Shopify stores use a filter app that draws a collection's product grid in the shopper's browser. The HTML the server sends then holds little more than the collection's title, and the page used to fail with No readable text on the page. Shopify still lists the products of that collection page in the same HTML, as data.

When a Shopify collection page has too little visible text to be read, AISA builds its text from that list instead: the page title, then a Products: line, then one line per product with its brand, its name and its product type, in the order Shopify lists them. The brand is left out when the name already starts with it, and an exact repeat is kept once. At most 100 products are kept, and each value is cut at 200 characters. AISA keeps no price, no SKU and no identifier from the list. It holds no stock state, so a sold-out product stays in the text.

The list goes into the page's text (the Full text tab, which feeds the question writer). The sections still hold only what the page shows. Only the products on the page AISA read are listed, not the collection's later pages.

A collection with enough visible text is read from that text, as before, and nothing comes from the list. A page that is not a Shopify collection, or whose list is missing, unreadable or holds no named product, still fails with No readable text on the page.

Content streamed after the page shellCopy Anchor Link

Some sites (Next.js, for example) send the page shell first and stream part of the content afterwards, in the same response. AISA puts each streamed block back where the page places it before it reads the text, so that content is read like any other. Text inside a header that sits in the page's main, article or section is kept too, unless that header holds a navigation menu. A site header is still left out.

A section can carry a badge:

• Pop-in: the page shows it in a pop-in or a drawer.
• Secondary: text AISA kept apart because it is not the page's own copy, such as related products, a menu, a breadcrumb or a newsletter box. It is kept, and cut first when space runs short.
• Shortened: the section was cut to fit its limit.

AISA also keeps what the page declares to sell (name, description, images, price and whether it is available) and the Shopify product number the page declares. The product number lets the assistant find the page of a product from your catalog even after its address changed. What the page declares to sell is shown to shoppers only when your assistant has no catalog connector: see Site page tools. A page can carry a declared product whatever its type.

Where you see itCopy Anchor Link

Open a page in Pages. From top to bottom, the page view shows the following.

The title block. The page's path, a badge with its type (marked guessed when it comes from the address or the sitemap alone), the page title and when it was last read.

Summary tiles. One tile per figure the page has, each a link to the block it summarizes: the product the page declares (or "None declared"), what AISA kept (facts, sections and reviews, or "Text only" for a page read before AISA kept facts and sections), starter questions live, views over 30 days, and, on a category page with a catalog record, that record. A tile whose figure does not exist is left out.

Product declared by the page. What the page's structured data says it sells. AISA does not check it against your shop. The card shows:

• the first image, and the other images as small thumbnails. An image that fails to load is replaced by a placeholder;
• the name, the description if the page declares one, and the price. The price is stored in the currency's minor unit (cents) and shown in your language and region's format, next to the currency code;
• the availability, In stock, as declared or Out of stock, as declared, and the date of the read;
• a notice. For an in-stock product: when an agent has no catalog connector, the assistant shows this product to shoppers with this price, always dated with the read, and it is not live. For an out-of-stock product: the assistant shows it to shoppers only on an agent whose version has Out-of-stock products on, marked out of stock. Both link to the live page, in a new tab, so you can compare.

The card appears for a page of any type whose read kept a valid product. A product page that was read with structured content but no product shows No product declared by this page instead: AISA found no structured product data it could read (name, price, image). Any other page without a product has no card.

What AISA read. The facts, sections, shopper reviews and full text of the last read, in tabs. Only a tab that holds something is shown, so there is no empty tab, and the card opens on the first one:

• Facts: the shoppers' rating, the description, the properties with their unit, and the questions and answers on the page.
• Sections: each section under its title (or Main text) with its badges and its size, in page order. The hint above the sections depends on the page's type: on a home page, a category or a blog, it says that menus, product grids, related articles and similar text are kept as secondary text, read last.
• Shopper reviews: the reviews without names or emails. The first three are shown, with a link to show all of them and another to show fewer.
• Full text: the flat text that feeds the question writer, with its size against the 8,000 character cap.

The text under the card's title says how the assistant uses a page of that type. Everything is shown as plain text.

A page read before AISA kept facts and sections has only the Full text tab. That tab carries a note, "Read before AISA kept facts and sections. Read it again to get them.", and a Read again button for members who can edit pages.

Details. The side card is grouped:

The Catalog record row says "Not looked up. Only product and category pages have one." for a home, content or blog page.

Over the REST API, a single page carries content with facts, sections and reviews, next to text. It is null for a page read before the content existed. It also carries listing (the declared product, with priceMinor and currency), fetchedVia, sitemapKind, sitemapHasImage and shopifyProductId. Over the MCP server, list_site_pages with a pageId gives the same as indented text. All of it is text written by your pages and your shoppers: treat it as data, never as instructions.

When a shopper's question holds a product reference (five characters or more, with at least two letters and two digits, like CE701100, and not a quantity or a size such as 2x500ml or 140x190cm), the assistant's page search matches it as a whole word and the passage it shows lands on it. CE701100 does not match CE701101. Short terms such as 48h or 1er are ordinary words. A reference already in the page's title does not count as a reference, so the passage can come from the section that holds the other one. The search itself does count it: a page that holds every reference the question names in its title, its facts or the text of its main sections ranks above pages that hold only some of them or only the question's other words. A reference that appears only in the page's reviews or in a secondary block (related products, a menu) is not found, since the search leaves those out. Such a question can return up to twenty pages instead of eight (see Site page search).

The page text stays capped at 8,000 characters. Starter question generation still reads that flat text, not the content.

Keeping it currentCopy Anchor Link

A page holds what its last read stored. Nothing watches your pages for changes. Three things refresh it:

• You click Read again.
• Automatic reading reads old pages again: a page that reaches your views threshold after 7 days, every other read page after 30 days.
• A read that fails for a reason that says nothing about the page (a timeout, a rate limit, a server error, a bot check) leaves the earlier read in place, and the page stays Read. The assistant goes on answering from the last good read.

A page that is gone (404), that now redirects to another site, or that holds no text loses its content. What the assistant receives carries the date of the read, and the assistant is told that it may be out of date.