The Agent Makes the Recommendation Now: From the On-site Engine to the Agent's Ranking

Who makes the recommendation now?

The agent makes the recommendation now: ChatGPT, Google AI Mode, Perplexity or Amazon Rufus takes the user's "find me something within budget, delivered in two days, easy to return", scans the catalogues, ranks three or four products, and your brand is either on that list or not. The classic product recommendation system worked inside the site: it processed behavioural data, produced the "customers who bought this also bought" block and grew the basket. That system is still needed; but in 2026 the first recommendation engine a buyer touches is not your site, it is the user's assistant. The on-site engine optimises the visitor; the agent's ranking decides whether the visitor arrives at all.

We first published this article in June 2024 to explain recommendation algorithms (collaborative filtering, content-based filtering, hybrid models) and measurement. Those sections remain in shortened form. The new layer is the criteria of the choice the agent makes on the user's behalf and how a brand enters that choice. Our thesis at Webtures: however good your on-site recommendation engine is, a product that never enters the agent's shortlist never sees that engine.

How does an agent choose a product?

An agent's selection logic is less a ranking algorithm than a triangulation. In our measurements and in industry studies four signal families stand out:

• Structured product data. Product, Offer and Review schema plus a complete product feed are the entry ticket. ChatGPT pulls most of its product data from Google Shopping-sourced feeds; a product not in a feed does not enter the comparison.
• Third-party authority. Reviews, independent "best X" lists, comparison articles, Reddit and expert media evaluations. The agent trusts independent sources that verify the brand, not the brand site; most product cards are fed from those sources.
• Transaction signals. Price, delivery time, return terms, stock and merchant reliability. Perplexity states its reasoning openly: it combines price comparison sites, expert reviews and user feedback. On Rufus, fulfilment speed and Prime eligibility correlate strongly on time-sensitive queries; listings with 15 or more answered Q&As appear in recommendations roughly three times as often.
• Accessibility. If OAI-SearchBot is blocked the catalogue is invisible however good it is; if the page loads via JavaScript the product is absent from the raw HTML.

The picture is close to human behaviour: a shopper reads reviews, compares prices, asks about delivery. The difference is scale and speed. The agent does this on every query, across dozens of sources, in seconds, and it does not say "I think"; it ranks. Industry measurements put ChatGPT's accuracy at 64% on standard product queries and 52% on multi-constraint queries; errors exist, and a large share of them come from the brand's own inconsistent data.

How do the on-site engine and the agent's ranking differ?

The two systems do not replace each other; they run in sequence. When the agent brings the user to your site (the dominant model in 2026 is "discover in AI, buy on site") the on-site engine takes over and grows the basket. But the first door belongs to the agent.

How do you get onto the agent's shortlist?

1. Give complete, consistent product data. Identity (GTIN/MPN), price, stock, delivery and returns identical on four layers; the detail is in product descriptions agents can read.
2. Accumulate independent evidence. Product reviews, expert reviews, a place in category lists, Reddit and community mentions. The large majority of brand mentions come from third-party pages; the "best" sentence you write on your own site is not evidence for an agent.
3. Make Q&A and returns visible. On Rufus, Q&A is the highest-leverage signal; Google's question_and_answer attribute and the MerchantReturnPolicy schema do the same job on the web. There is no answer to the agent's "is it easy to return" without data.
4. Make delivery and merchant reliability measurable. Country-level delivery time, shipping cost, seller rating and review count as numeric fields in the feed; the agent matches them against constraints.
5. Join the platforms' merchant programmes. The OpenAI product feed, Google Merchant Center, the Perplexity Merchant Program (zero commission), Prime eligibility on Amazon. An unregistered brand appears on most surfaces only indirectly, through third-party lists.
6. Keep access open. Allow the search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot) apart from any training-bot decision; render on the server.
7. Measure visibility. On a defined prompt set, track in how many answers, at which position and with what reasoning the brand is recommended; check whether competitors are the answer in your place. We do this on five surfaces with Brantial.

What changes for the on-site engine in the agent era?

The algorithm families are the same: collaborative filtering learns from users with similar behaviour, content-based filtering from product attributes, hybrid models combine the two, and a session-based layer reads the live click sequence. The measurement principle is the same too: success is measured not by click-through but by incremental revenue against a control group. Three things change.

First, the entry signal. A visitor arriving from an agent channel usually comes with constraints already set; a user who arrived with "linen shirt, size 42, in two days" should see complements that fit the constraint rather than a random "bought together" block. Second, a new answer to the cold-start problem: the product attributes you use for the content-based score are already structured in the feed; the same data feeds both engines. Third, memory. Agents remember user preferences across sessions; a brand that fails its delivery promise on the first experience does not make the list on the next query. The on-site engine's job is to keep, on the site, the promise the agent made.

What are the most common mistakes?

• Treating the agent as a new ad surface. You do not pay your way onto the shortlist; you enter with data, evidence and transaction signals. Advertising is a separate layer (ChatGPT Ads, Google AI Max) with limited access from some markets.
• Counting self-praise as evidence. "The country's most preferred brand" is an unverifiable claim for an agent; 412 reviews and a 4.6 rating are data.
• Hiding delivery and returns on a policy page. A return window the agent cannot read from the product page is treated as unknown.
• Focusing on one surface. A brand that enters ChatGPT but is absent from Google AI Mode misses a large share of users in many markets; five surfaces are measured together.
• Confusing visibility with traffic. If the agent recommends you but sends no referrer, GA4 shows "Direct"; whether you are on the shortlist is told by prompt measurement, not analytics.

How does the Webtures approach work?

We measure whether a brand enters the agent's recommendation in three steps: first, visibility on five surfaces with a prompt set defined for the target market and language (prompt coverage, found rate, average position, citation rate, share of model); then an analysis of which signal puts the recommended competitors ahead (review count, delivery promise, price band, third-party lists); finally, prioritising the gaps in the brand's data and evidence layer. In our own export matrix this analysis focuses on the product feed and the independent review ecosystem in B2C markets such as the United States and the United Kingdom, and on the readability of supplier data on the agent exchange in B2B markets such as Germany.

Read your recommendation engine's place in the agent era together with inventory truth and category architecture; we covered how return and trust signals affect agent visibility in a separate article. To see in which prompts your brand is recommended, start from the Agentic Commerce Readiness page and set up the measurement set with our Agent Experience team.