A deal tracker shows how to make prices readable and citable by AI assistants

A build log from a small deal tracker

As dev.to reported, the team behind Savlume, a deal tracker for the US, has published the second post in a series about how the site works. The first covered its deal verifier and the layer of files and endpoints built for AI agents. This one focuses on two daily jobs: tracking prices automatically, and making the site something AI search tools can cite.

It is a small, builder-written post, and the claims are the team's own. Still, the design choices are specific enough to be useful beyond this one site.

The record matters more than the scraper

The authors argue that the question people ask most, what they scrape, is less important than what they save. Each deal is a record with several parts: the price, a verification status, a verification timestamp, the conditions attached, and a reference price held apart from the checkout price.

The status has two values. Retailer-verified means the price was confirmed on the retailer's own listing. Source-reported means it came from a deal source and has not yet been confirmed at the retailer. Conditions cover things like whether a coupon or Prime membership is needed, whether there is a per-customer limit, and whether shipping is free.

The team's reasoning is blunt: a price with no timestamp is a rumor, and a discount measured against an inflated reference price is misleading. Keeping the two prices separate stops the second problem at the source.

Rules the system won't bend

Two rules are described as never broken. Expired or out-of-stock entries stay in the database as reference but are never offered as recommendations. And money you might get back later, such as cashback, gift cards or loyalty points, is not used to lower the price shown at checkout.

Price history follows a similar discipline. Comparisons only line up the same product, the same variant and the same conditions. When a piece of information is missing, it stays marked as unknown rather than being filled in with a guess.

• Stale deals are kept for context, not advice.
• Rebates don't change the headline price.
• Unknown stays unknown.

Making a site that AI can quote

The second half covers GEO (generative engine optimization) and AEO (answer engine optimization). The post treats them as one idea: when someone asks ChatGPT, Perplexity or Gemini for the best air fryer under $100, the answer is built from sources the model can read and cite. An ordinary web page is awkward for an agent to quote exactly.

So Savlume publishes a machine-readable layer alongside its pages:

• /llms.txt, a single index of the site: methodology, today's deals, coupons, categories, stores, brands, reviews and all 35 guides. It also lists rules telling agents to quote the original source and the canonical Savlume item link, with publication, update and verification times.
• /agent, an API guide written for agents rather than people.
• /api/mcp, an endpoint that lets an agent query the catalog directly.
• Per-item Markdown, a clean text version of every item, which is what an agent actually wants to quote.

The catalog is also available as paginated JSON, as an RSS feed for all updates, and as a JSON record for each item. The 35 guides under /best combine review consensus from several sources with current retailer offers, so a tracker can follow a whole category rather than one product.

Three lessons and one open question

The team lists three takeaways. First, write the llms.txt early rather than last, because drafting it pushes you to say what the site is and what proof backs each claim. Second, put uncertainty inside the data, so the words "source-reported" sit right next to the price instead of in a footnote. Third, keep the canonical link in every answer, since that is how a citation survives.

The payoff, in their telling, is that an agent can give a real price, a real timestamp and a real link instead of inventing one, and can say how well the price is confirmed. For the site, each AI answer that cites a Savlume price becomes a timestamped backlink.

What the post does not show is evidence that this extra layer wins more citations than a well-built ordinary page would. That remains an open question, so treat the approach as a sensible design, not a proven tactic.