What Changed in CitePulse in September 2026: Weekly Monitoring, Your Own Questions, AI Bot Logs and Draft Studio

One day of changes, all live: every monitoring plan now measures weekly, you add your own buying questions on top of the audit series, the dashboard shows which pages the AI crawlers actually read, Draft Studio assembles structural blocks from your approved facts without a language model, and the question × engine matrix says plainly why Microsoft Copilot is not measured. Here is what each one does, what it deliberately does not do, and what it costs.

Weekly monitoring on every plan

Shield used to measure once a month. Cited sources turn over by roughly two thirds day to day, so a monthly snapshot told you very little about the direction of travel. From today Shield, Protect and Scale all run weekly. Each run asks the frozen five questions from your audit, three times per engine, so the runs stay comparable and “seen in k of n runs” keeps its meaning. The price did not change.

Your own buying questions

The five audit questions are grounded in your own site and frozen as a series, which is what makes week-over-week comparison honest. Most tools count “prompts” instead, and 25 to 100 of them at entry. The two are now compatible: on a monitoring plan you add your own buying questions on top of the series, up to 10 on Shield, 20 on Protect and 30 per client domain on Scale. Every run asks them alongside the series, and the dashboard reports a question added later only against the runs that actually asked it, never as “missing” in earlier runs.

If you want one number to compare with a prompt count, use measured answers per month: questions × engines × three repetitions × weekly runs. That is roughly 580 on Shield with ten own questions, about 1,080 on Protect and about 1,500 per client domain on Scale. A question asked once is a coin toss; the repetition is the point.

AI bot logs: which pages the crawlers read

Being cited starts with being read. The new AI bots tab shows requests from GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Googlebot, Bingbot, Applebot and a dozen others, with why each one reads: a training corpus, a live search index, or a fetch while answering a user. Two ways in: connect your Cloudflare zone with a read-only analytics token (verified on connect, stored encrypted, removable in one click), or upload a log file in nginx, Apache or Cloudflare Logpush format. Only day, crawler and path aggregates are kept. Raw lines and IP addresses are discarded while parsing and nothing raw is written to storage.

The part nobody else does at this price is the crossing with your cited pages: read and cited, read but never cited (the content, not the access, is the question), and cited but rarely read (a quote served from an index, so freshness may lag). The tab also puts robots.txt next to reality, and counts how many times a crawler fetched llms.txt, which is usually a very small number and is exactly why we keep calling llms.txt agent-readiness hygiene rather than a citation factor. When a crawler that used to visit goes quiet for 14 days, you get an observation, not a diagnosis. Shield shows totals per crawler; Protect and Scale show the pages and the crossing. Cloudflare analytics are sampled, and the tab says so.

Draft Studio: structural blocks, no language model

CitePulse does not write articles or marketing copy, and that has not changed. What changed is a narrow, deterministic layer for the questions where your own page can win (door A). Draft Studio assembles three blocks: an FAQ block with FAQPage schema, with the questions we already measure and answers composed only from your approved fact registry; a comparison-table skeleton with your column filled from facts and the competitor cells left empty for you to fill from their public pages; and a visible last-updated line with dateModified.

No model runs. Every sentence traces to a fact id. Where no approved fact covers a question, the block shows a visible gap instead of a guess, and a guard refuses any number that does not come from a fact or the date. You download the block, paste it, and mark it done; the Action Center then compares the following runs against the engines’ own noise band. A short model-written answer paragraph, with a verifier and your explicit approval, is a separate decision we will take after the doors A/B data has accumulated. It is available on every plan and needs at least one approved fact.

Copilot: not measured, and we say why

Microsoft offers no official API for the consumer Copilot. The documented Copilot APIs cover Microsoft 365 Copilot inside a licensed tenant, and Azure’s Grounding with Bing Search generates our own model’s answer rather than showing what Copilot says. Tools that claim to track Copilot automate a browser session, which produces results that cannot be repeated or audited. The matrix now carries a Copilot column marked “not measured” with that reason, and we will add it the day official programmatic access exists.

SEO × AI: Search Console and GA4

Already in the dashboard, now also on the homepage where it belongs: connect Google Search Console to see where Google ranks a page that AI does not cite, and the reverse; connect GA4 to see the sessions and conversions that already arrive from ChatGPT, Perplexity, Gemini and Claude. Both connections are read-only and can be disconnected at any time.

What did not change

The free audit is the same five-question, three-engine, three-repetition matrix with a fixed budget. Google AI Overviews are read as a source, not counted as an engine. The public noise panel keeps running every day. And every result is still an observation: after this change, beyond or within the band. Not proof of cause.

Frequently asked questions

How often does CitePulse monitor now?

Weekly, on every plan: Shield, Protect and Scale. Each run asks the frozen five questions from your audit plus your own questions, three times per engine.

How many of my own buying questions can I add?

Up to 10 on Shield, 20 on Protect and 30 per client domain on Scale, on top of the five audit questions. A question added later is reported only against the runs that asked it.

Does CitePulse store my server logs or IP addresses?

No. Only day, crawler and path aggregates are kept. Raw lines and IP addresses are discarded while parsing, and nothing raw is written to storage.

Does Draft Studio use an AI model to write?

No. It assembles structural blocks from your approved fact registry with deterministic templates. A missing fact becomes a visible gap, never a guess. A short model-written answer paragraph with a verifier and your approval is a separate, later decision.

Why is Microsoft Copilot shown as not measured?

Microsoft offers no official API for the consumer Copilot. Tools that claim to track it automate a browser session, which cannot be repeated or audited. CitePulse marks the column as not measured and will add it the day official access exists.

See it before you sign in: the live demo runs on sample data for a fictional company, or run your free audit. Related reading: how much AI answers change from day to day and what llms.txt is and who reads it.