# How to Structure a SaaS Product Page So AI Assistants Can Recommend It > Source: > Published: 2026-05-21 07:42:32+00:00 Most SaaS product pages in 2026 are structured for human eyes and Google's old keyword crawler. They render fine in a browser, they score well on Lighthouse, they have a /blog and a /pricing and a /about . And they are nearly invisible to ChatGPT, Claude, Perplexity, and Gemini when a user asks "what is the best tool for X." This post is the technical playbook for closing that gap. PeerPush, the indie SaaS directory at https://peerpush.net, builds every product listing around the same primitives covered below. The advice applies whether a product gets listed on PeerPush or not. The structure is what matters. The TL;DR: AI assistants synthesizing a "best X for Y" answer fetch pages, parse facts off them, and rank candidates by the volume and quality of extractable structured data. Make your facts extractable, on stable URLs, with the right schema.org annotations, and the assistant has what it needs to recommend you. A frontier AI assistant fielding a product recommendation question runs roughly this pipeline: 1. Retrieve: fetch pages relevant to the user's query 2. Extract: pull discrete facts (name, price, features, integrations, audience, tradeoffs) 3. Synthesize: build a ranked shortlist and a justification paragraph The extraction step is where most product pages fail. A page that says "We help your team scale operations with our intuitive platform" is invisible. A page that says "Acme Tasks is a project management tool for engineering teams of 3-30. Free plan with 5 users. Paid plans start at $12/user/month. Integrates with GitHub, Linear, Slack" is the page the assistant cites. Structured data formats give the model explicit hooks for the extraction step. The most important ones for SaaS are JSON-LD with schema.org's SoftwareApplication , Offer , FAQPage , and BreadcrumbList types. The minimum viable JSON-LD for a SaaS product page in 2026: { "@context": "https://schema.org", "@type": "SoftwareApplication", "name": "Acme Tasks", "applicationCategory": "BusinessApplication", "operatingSystem": "Web, iOS, Android", "description": "Project management for engineering teams of 3-30. Async-first, GitHub-integrated, opinionated about cycle time.", "url": "https://acmetasks.example/", "offers": [ { "@type": "Offer", "name": "Free", "price": "0", "priceCurrency": "USD", "description": "Up to 5 users, unlimited tasks, no SSO" }, { "@type": "Offer", "name": "Pro", "price": "12", "priceCurrency": "USD", "description": "Per user / month, unlimited users, SSO, audit log" } ] } Two things matter here that most product pages get wrong: offers must contain price and priceCurrency as primitive values, not as priceRange: "$12-$48" . AI extractors and Google's rich result parsers both reject priceRange on Offer (it's only valid on LocalBusiness ). Multiple price points = multiple Offer entries in an array.description does double duty. It is what the assistant pulls when summarizing your product. Write it as a single noun-phrase sentence that names who the product is for, not as marketing copy.A common mistake: putting the @type as Product instead of SoftwareApplication . Product is for physical goods. SaaS is software. Google and the major AI extractors weight SoftwareApplication correctly; Product confuses the classification. If your product page lives at /products/acme-tasks under a category tree like /categories/project-management , give the assistant the path: { "@context": "https://schema.org", "@type": "BreadcrumbList", "itemListElement": [ { "@type": "ListItem", "position": 1, "item": { "@id": "https://acmetasks.example/categories/project-management", "name": "Project Management" } }, { "@type": "ListItem", "position": 2, "item": { "@id": "https://acmetasks.example/products/acme-tasks", "name": "Acme Tasks" } } ] } Note: item must be a full Thing object with @id and name , not a raw URL string. Google's documentation shows both shapes; some AI extractors only accept the strong-typed form. Use the strong-typed form. When a user asks "does Acme Tasks support SSO," the assistant wants a Q&A pair it can pull verbatim. The FAQPage schema gives it one: { "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Does Acme Tasks support SSO?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. SAML SSO and SCIM provisioning are available on the Pro plan and above. Free plan users authenticate with email + password or Google OAuth." } }, { "@type": "Question", "name": "What is Acme Tasks priced at?", "acceptedAnswer": { "@type": "Answer", "text": "Free for up to 5 users. Pro is $12 per user per month, billed annually. Enterprise is custom." } } ] } Five to ten Q&A pairs per product page is the sweet spot. Cover pricing, integrations, the most common "does it support X" questions, and at least one objection ("how does this compare to [competitor]"). llms.txt is the AI-era equivalent of robots.txt . It is a plain Markdown file at the root of your domain (https://yourdomain.example/llms.txt ) that gives AI crawlers and assistants a curated index of your most important public URLs. A minimum useful llms.txt for a SaaS product: # Acme Tasks > Project management for engineering teams of 3-30. Async-first, GitHub-integrated, opinionated about cycle time. ## Core pages - [Pricing](https://acmetasks.example/pricing): full plan breakdown with real prices - [Integrations](https://acmetasks.example/integrations): supported and not-supported integrations - [Documentation](https://acmetasks.example/docs): how the product works - [Changelog](https://acmetasks.example/changelog): what's shipped recently ## Comparisons - [Acme Tasks vs Linear](https://acmetasks.example/alternatives/linear): when each is the better choice - [Acme Tasks vs Jira](https://acmetasks.example/alternatives/jira): tradeoffs and migration notes ## Posts AI assistants might find useful - [How we think about cycle time](https://acmetasks.example/blog/cycle-time): opinion piece, 2026-01-15 This file is not a standard yet (the proposal is at llmstxt.org ). It works in practice because AI crawlers and grounded retrieval systems are increasingly opportunistic about consuming Markdown indexes when they find them. Cost to ship: 30 minutes once, then keep current. Upside: a small structured nudge that tells AI infrastructure which URLs matter most. The compounding effect of being in AI training corpora rests on URL stability. Every URL killed is a citation killed. Three rules that pay back over months: Never just delete a page. Use 301 redirects when restructuring. If a feature is sunset, leave the page up with a "this feature is no longer supported, here is the current alternative" notice and link to the replacement. The URL is what AI engines remember; the content can change underneath it. Canonical tags on every product page, pointing to the version you want indexed. Especially important if you syndicate content to dev.to, Hashnode, or Medium (which is exactly what this post is doing, by the way: the canonical of this article is https://peerpush.net/blog/2026-indie-saas-launch-playbook, declared in the canonical_url frontmatter dev.to reads). sitemap.xml discipline. Submit it via Google Search Console. Update it on schedule. AI training pipelines often hit sitemap.xml as a seed for discovering a site's structure; an up-to-date sitemap with a per URL signals which pages are current. A practical loop for testing whether AI assistants are picking up your changes: Pick 5-10 questions a real user might ask in your category. Examples for a project management tool: "What is the best project management tool for engineering teams?" "What are alternatives to Linear?" "Does Jira have a free plan?" Run each question through ChatGPT (paid), Claude (free), Perplexity (free), and Gemini (free). Record whether your product is mentioned, at what position in the answer, and which competitors are mentioned alongside. Repeat the same prompts weekly. Single-week deltas are noise. The trend over 6-12 weeks tells you whether the structured data work is moving the needle. PeerPush runs a daily version of this loop internally to measure AI citation rate. The same loop works for any product team that cares about being recommended. description written as marketing copy. Use noun phrases that name who the product is for and what it does, not adjective phrases that describe how it feels.search.google.com/test/rich-results ) catches schema errors. If it doesn't pass there, AI extractors built on similar parsers won't pass either.robots.txt to "protect content from AI training." That choice trades short-term content protection for long-term invisibility in AI recommendations. Worth thinking about explicitly rather than blocking by default.A practical checklist, in priority order: /pricing page. Are there real numbers, real plan names, real trial terms? If not, fix that first.SoftwareApplication + Offer to the homepage and the pricing page./alternatives-to- page with the comparison-table pattern from earlier in this post.FAQPage JSON-LD block with 5-10 questions on the homepage or a dedicated FAQ page.llms.txt at the root of your domain with the format shown above.Most of this is implementable in a single afternoon. None of it depends on a viral launch day. All of it compounds in AI training corpora over the next 24+ months. The new shelf is permanent. The structured data is the door key.