Auto page-aware questions

Auto page-aware questions make the widget feel smart on every page with zero setup. Instead of hand-writing starter questions for each page, the agent reads the content it already indexed for a page and drafts a few suggested questions a visitor might click — product questions on a product page, category questions on a category page, your default starters everywhere else.

It is on by default. Turn it off per agent under /app/agents/{agent}/settings.

How it works

• When a page is crawled and indexed, a background job drafts 3–4 short questions grounded in that page's own knowledge-base chunks. The model is explicitly told to propose only questions the page content can answer — so the widget never suggests a question the bot then can't answer.
• The questions are stored per (page, language). When the page is re-indexed (its content changed, or you click Reindex), they refresh automatically.
• On widget load the questions are a single indexed lookup — no AI call on the visitor's hot path, so first-token latency is unaffected.
• Generation happens for the languages your workspace has enabled, so a Dutch visitor and an English visitor each see questions in their own language.

How suggested questions are chosen

The widget picks starter questions in this order, first match wins:

1. A Smart Segment's configured starters. If the page resolves to a segment that has starter prompts set, those always win — the segment is the manual override.
2. Auto page-aware questions generated for this exact page and language (from the crawled content — see above).
3. Instant product questions read live from the page's own product data — see below. Kicks in on a product page that hasn't been crawled yet, so even the first visitor gets product-specific questions.
4. The agent's default starter prompts (or the vertical preset's) when none of the above applies.

So segments remain the escape hatch when you want exact control over a page, while every other page still gets smart, page-specific questions for free.

Instant product questions (no crawl needed)

Auto page-aware questions (tier 2 above) need the page to be crawled first: the very first visitor to a brand-new product page would still see generic starters until the page is indexed. On a webshop with thousands of products that gap matters, so there is a faster path that needs no crawl at all.

Every modern product page ships Schema.org structured data — the same JSON-LD Google reads for rich snippets, and the same data Shopify, WooCommerce and Magento already emit. The widget reads it directly in the visitor's browser and turns it into grounded starter questions, filling in the product name:

• Tell me more about {product} — whenever a product name is present.
• What does {product} cost? — only if the page exposes a price.
• Is {product} available? — only if the page exposes availability.
• What features does {product} have? — only if the page has a description.

Each question is only offered when the backing fact is on the page, so the assistant can always answer it from the page context it already receives with every message — never a question it then can't answer. The wording is fully translatable (the lang/{locale}.json keys "What does :name cost?" and friends), so a workspace can re-word or re-tone the templates without any code. Product names are filled in client-side and never stored — nothing a visitor's page claims is ever written to the knowledge base.

This runs only when the agent has auto page-aware questions enabled and no higher-priority starters won. Once real visitors have browsed a product page, auto-indexing crawls it and tier 2 (crawled, AI-drafted questions) takes over automatically.

See it live

/demo/product is a self-contained sample product page with real Schema.org data and the widget embedded. Append ?agent=YOUR_AGENT_ID to load a specific agent (its allowed origins must include this site), and ?locale=nl to see the questions in Dutch. Open it and the assistant suggests questions about that exact product on the first visit.

Regenerate on demand

Agent settings has a Regenerate button that re-drafts the questions for every indexed page of the agent — useful right after you add sources or change a lot of content and don't want to wait for the next re-index. From the CLI:

php artisan pitchbar:generate-page-questions {agent-id}
php artisan pitchbar:generate-page-questions --all   # every enabled agent

Both queue one generation job per distinct page URL on the index queue. Existing pages have no generated questions until they are re-indexed or regenerated, so run this once after enabling the feature on an established agent.

When a page produces nothing

If a page's indexed content is too thin to support good questions, the generator returns nothing for that page and the widget falls back to the agent's default starters — never a blank or a low-quality guess.