The AI analyst
How the analyst plans, investigates with tools, learns from your decisions and stays within budget.
Agent, inside firm limits
After the deterministic part of an audit, an AI analyst takes over. It is an agent: it decides which steps are worth running, chooses which tools to call, and checks its own answer. It runs inside fixed limits on steps, AI calls and budget, and every step lands in the audit's Activity log with its reasoning and cost.
The analyst only reads. Its tools can't change your site, and its output is recommendations that wait for you.
Steps
| Step | What it does | Runs when |
|---|---|---|
| Sample | Picks up to 50 of the most important pages (inbound links, content, titles). | Always |
| Plan | Decides which of the next steps to run, with a reason for each. No AI call. | Always |
| Content review | Reviews up to 20 sampled pages' content quality. | Sampled pages have content |
| Search intent | Compares what searchers want with what each page offers; can add gsc.intent_mismatch. | Search Console has 28-day query data |
| Investigate | Calls read-only tools in bounded rounds and writes focus areas with evidence. | Any findings: 3 rounds for medium/low, 6 for critical/high |
| Synthesize | Writes diagnoses and recommendations, each citing evidence. | Always |
| Revise | If the plan repeats advice you rejected, revises it once. | A repeat is found and budget remains |
Tools
During the investigation, and when you [chat](/docs/chat-and-goals), the analyst can call:
| Tool | Returns |
|---|---|
get_pages | Crawled pages, filtered and summarised |
get_page_detail | One page: title, description, headings, canonical, robots, structured data |
get_findings | The audit's findings by check or severity |
get_links | Internal links to and from pages |
get_pagespeed | Core Web Vitals and PageSpeed results |
get_issues | Tracked issues and their history |
get_gsc_data | Search Console queries and pages (when connected) |
Results are size-limited so the analyst sees summaries, not dumps, and recommendations may only cite evidence that exists.
Memory and goals
Before writing advice, the analyst reads the site's history: up to 30 recent decisions, grouped as rejected (with your reasons), worked (outcome improving) and no effect yet (flat or worse). The rules it follows: don't repeat rejected advice unless the evidence changed, build on what worked, and change approach where nothing moved.
It also reads the site's active [goals](/docs/chat-and-goals): it investigates what stands between the site and each goal first, ranks advice that moves a goal higher, and says which goal a recommendation serves.
Cost and limits
- Per-audit ceiling: at most 40 AI calls per audit (60 on your own key). Calls are reserved so the final steps always have room.
- Monthly budget: owners set it in Settings → AI. No new AI call starts once the month's spend reaches it.
- Cache: an identical request within 30 days is answered from cache, free.
- Ledger: AI usage in the sidebar lists every call, model, token count and cost, and whose key paid.
ai.skipped_budget notice. If a limit is hit mid-run, the AI output is marked degraded. Either way the technical audit, score and issues are complete.Models
Steps use three tiers: a cheap model for page reviews and intent, a medium model for the investigation, and a strong model for synthesis and drafts. The platform provides models by default; you can use your own provider and key, see [AI providers and budget](/docs/ai-providers).