AI for competitor prices

The AI for Competitor Prices module adds the AI to the Competitor Prices app. The AI does the slow parts of setting up price tracking:

  • it analyses a competitor website: shop engine, search page, sitemap;

  • it finds your products on the competitor websites and proposes the matching pages;

  • it builds an extraction rule for pages whose price cannot be read from the structured data;

  • it repairs a broken listing by writing a new rule when a website changes;

  • optionally, it reads the price itself when no rule works.

The AI never sets a price and never confirms a doubtful match: a person confirms every proposed listing, and sales prices change only when a manager accepts a recommendation.

Note

The module requires the AI application with at least one chat model set up (see Providers and models) and the Competitor Prices app. The AI calls run with the access rights and limits of the user who clicks, or, in the background, of the user chosen in the settings.

Warning

To judge a page, the AI provider receives the text of the competitor’s product page and the name, codes and brand of your product. Every call uses your AI quota. Content of a web page is treated as untrusted data: instructions written in a page are not followed.

Settings

Go to Competitor Prices ‣ Configuration ‣ Settings, section AI.

Setting

Meaning

AI Models

One model per task. Rules and finding writes and repairs extraction rules, picks links from search results and finds a shop’s search page. Product matching decides whether a competitor page sells the same product (model, variant, pack size) and reads prices when no rule works: use the most careful model here. Empty: the main AI model.

Scheduled AI Work Runs As

The user whose AI rights and limits are used for the background work (product finding and repair). Empty: the AI does no background work.

Confirm Matches with the Same GTIN

A found page whose GTIN (EAN) equals the barcode of the product is confirmed without review. Every other match waits for a person. On by default.

Repair Broken Listings with AI

When a listing becomes broken, the AI writes a new rule and checks it before use, at most AI Repairs per Website per Day (5) times per website.

AI Reads the Price When No Rule Works

Costly: one AI call per page that no rule can read. The price is marked as read by AI (method AI reading).

screenshot: ai-competitor-prices-settings
menu
Competitor Prices ‣ Configuration ‣ Settings
shows
The "AI" block of the Competitor Prices settings with the two model fields, the "Scheduled AI Work Runs As" user, the GTIN and repair options and the "AI Reads the Price When No Rule Works" option.
highlight
The "AI Models" setting (red frame).
data
A cheap model for rules and finding, a careful model for matching, an AI user selected.
notes
English UI, light theme, 1440px width, crop to the AI block.

Analyse a website

On the form of a competitor website click Analyse Website (managers). The app reads the home page and reports in a notification and in the chatter: the shop engine (Platform), the search page, the sitemap listed in robots.txt, whether robots.txt allows reading, and whether the home page carries structured product data. The search page and the platform are saved on the website (Search URL, Finding Products tab).

The Finding Products tab also shows the state of the sitemap (Sitemap: Not loaded, Loading, Loaded or Not available), the Addresses in Sitemap and Sitemap Loaded. The sitemap is loaded the first time it is needed and refreshed from time to time. On the Fetching tab, Render JavaScript With selects a web provider that draws the page for shops that show their prices only after a script ran.

Find products on competitor websites

Select products in the product list (or open one), then choose Action ‣ Find on Competitor Websites. On the product form the button is also on the Competitor Prices tab.

screenshot: ai-competitor-prices-discovery-wizard
menu
Sales ‣ Products ‣ Products ‣ Action ‣ Find on Competitor Websites
shows
The Find on Competitor Websites window with the explanation, the products, the competitor websites, "Skip Products Already Listed" and "Start Now", and the Find Products button.
highlight
The "Skip Products Already Listed" and "Start Now" options (red frame).
data
Five selected products, three competitor websites.
notes
English UI, light theme, dialog only.
Products and Competitor Websites

The products to look up and the websites to search (the websites with fetching on are preselected).

Skip Products Already Listed

Skips a product that already has a listing on the website.

Start Now

Processes as many as the request limits of the websites allow at once (about 40 seconds); the scheduled job (Competitor Prices: find products on competitor websites (AI), every 10 minutes) finishes the rest.

Click Find Products. For each product and website a search task is created. A task looks for the product with its GTIN (barcode), manufacturer part number, internal reference and name:

  1. on the search page of the website;

  2. if nothing fits, in the website’s sitemap;

  3. if nothing fits, with a web search engine, when one is configured in the AI settings and allowed.

Candidate pages are downloaded politely, within the limits of the website, and the page’s identity is compared with your product. A page with the same GTIN is confirmed at once (if Confirm Matches with the Same GTIN is on). A page with the same manufacturer part number and brand, or one the AI judges to be the same product (also checking the model, variant and pack size, which sets Units per Listing), is proposed: it becomes a listing To Review.

The tasks are in Competitor Prices ‣ Listings ‣ Product Finding:

Status

Meaning

Waiting

The task is queued or in progress.

Found (same GTIN)

A listing was created and confirmed.

Found, to review

A listing was proposed; confirm or reject it in Competitor Prices ‣ Listings ‣ To Review.

Not found

No matching page; the notes say what was tried.

Error

For example the website blocks automated access.

The filters Waiting, Found and Not Found narrow the list; Retry (on a not found or failed task) searches again, and the list action Process Now runs the selected tasks at once. If a website blocks the search, its fetching pauses as usual (see Competitor websites and fetch).

screenshot: ai-competitor-prices-product-finding
menu
Competitor Prices ‣ Listings ‣ Product Finding
shows
The Product Finding list with products, websites, statuses Found (same GTIN), Found, to review and Not found, the listing and the notes.
highlight
The status column (red frame).
data
About 15 tasks with all statuses.
notes
English UI, light theme, 1440px width.

Build a rule with AI

For a listing whose price cannot be read from the structured data of the page, click Build Rule with AI on the listing form. The AI:

  1. checks whether the page has structured product data (then no rule is needed);

  2. checks whether the website’s rule already works on the page;

  3. otherwise asks the model for a rule, audits it (the selector must match, point to a number and read the price the page shows), and, when other confirmed pages of the website exist, checks that the rule also works there.

A rule that works on several pages is saved as the website’s rule (Validated); otherwise it is saved as the listing’s Own rule, with Rule Source AI and a note. The price read is recorded at once. A rule that cannot be checked is not saved, and the message says why. You can read and edit the rule on the website or listing form, and Test Read shows what it reads.

Repair broken listings

When a listing becomes broken and Repair Broken Listings with AI is on, the listing is queued (AI Repair Queued). A scheduled job (Competitor Prices: repair broken listings (AI), hourly) builds and audits a new rule as above and, if it works, confirms the listing again. A failed repair is noted on the listing (AI Notes) and not repeated. The number of repairs a website had today is shown as AI Repairs Today on its Extraction Rule tab; the daily cap keeps the cost and the load on the website in check.

Tip

Start with a few products and one website, check the proposed listings and the rules, and only then run large batches.