Write for content gaps
Drafts posts for prompts and searches you lose. Of all the automations, this one adds to your AI visibility most directly. Each post answers one question engines answer without you, which also makes it easy to measure whether the post worked.
Autopilot collects three kinds of gaps and drops the ones a post wouldn’t help:
- Prompt gaps: prompts where competitors show up and you don’t. Dropped when you already win the prompt, when no competitor shows up either, when you’re already in more than half of the answers, or when a GEO brief exists for it.
- AI search gaps: searches AI engines ran while answering your prompts, without finding you. Same rules, and it needs a tracked prompt to measure on.
- Search gaps: Search Console queries with at least 100 impressions that your content doesn’t cover well. If Search Console already shows a relevant blog post for the query, Autopilot can refresh that post instead of drafting a new one.
Blog: … in your blog folder that answers the prompt or query directly, in your brand’s language and voice. Facts come from your brand, your sitemap and the evidence, and recent pull requests about the topic count as product facts. The pull request lays out the evidence, for example “AI engines answer … without you on 3 of 4 engines (12% mention you), while … show up”. Notra adds a generated hero image and checks every link the post adds. When the gap’s data says an existing page should be updated instead, and Autopilot finds that page’s file in your repository, it updates the page (Update: …).
A closed post can come back after 90 days if the gap is still open. Ideas from this automation expire early once their gap closes.
Refresh a blog post from Google Search Console
Connect Google Search Console, select the property in the same GEO project Autopilot uses, and connect the website repository. When a search gap recommends updating an existing post, Autopilot looks for one matching Markdown or MDX file. If it can’t resolve the file, it skips the refresh instead of drafting a duplicate. The pull request keeps the existing URL and frontmatter and edits only the article body. The edit answers the reader’s question directly, with details the evidence supports. After merge, Autopilot checks the original URL for new text from the draft. It allows up to 30 days to see that text; if you rewrite it during review, Autopilot might not recognize the update. It compares clicks, impressions, and average position for the same URL and selected searches: 14 full days before merge and 14 full days after it first sees the new text. Autopilot waits for finalized Search Console data and stops checking for a comparison 60 days after merge. If Google returns no matching data for a complete window, the report shows zero reported clicks and impressions, with no average position. The comparison uses Google’s overall Web search data. That can include AI Overviews and AI Mode, but Autopilot doesn’t separate their impressions or count AI citations here. Use the Generative AI performance report in Search Console for Google’s AI-specific impressions. A before-and-after comparison doesn’t prove the edit caused the change. Without GEO, the log says “GEO is not set up, so there are no content gaps to read”. When the planner holds off, it says why, like “Waiting: …” or “Not acting on … yet”.Refresh pages AI stopped citing
Updates pages AI cites less than last month, and posts Autopilot wrote that didn’t move the number.
A page qualifies when it was cited in at least two answers before, its prompts got at least three answers since, and its citation rate fell by at least half, by more than chance explains on that many answers. Posts Autopilot wrote for a prompt (from Write for content gaps or Write comparison pages) also qualify once, 45 to 180 days after the merge, when no AI answer has cited the post and no AI assistant has fetched it. Lost citations come first.
To update the page, it reads the pages engines cite instead (up to three) and your own pricing and plan pages for current figures, then goes through the page section by section: keep, update, add or cut. It makes at most 12 edits. It refuses a rewrite (more than 60% of lines) and an edit too small to matter, never edits dates (Notra sets the updated date in frontmatter), drops edits with dead links and follows redirects to their final address. It never invents numbers, customers, quotes or links. A figure the edits add that no page it read contains is flagged in the pull request, and so is a price or percentage that only a competitor’s page states. A flagged pull request never publishes on its own.
The pull request says what changed and why, for example “Cited in 40% of AI answers to its prompts a month ago, 12% now.”
Typical log lines when nothing is refreshed are “No page lost AI citations”, “No dropping page has a Markdown source”, “The page already answers what the engines look for” or “A page refresh is still waiting for review”.
Write comparison pages
Drafts comparison pages where AI picks a competitor. Buyers ask AI “X vs Y” and “alternatives to X” right before they decide, so these answers are worth a lot.
A comparison question asks for “vs”, “versus”, “alternatives”, “compare”, “better than”, “instead of”, “switch from”, “best” or “top”. A competitor qualifies when AI names it instead of you on at least two different comparison questions, or on one question that names the competitor itself. Autopilot picks the competitor that costs you most, skips ignored competitors and ones you already have a comparison page for, and writes an alternatives guide when a question names the competitor and asks for alternatives, a replacement or switching from it, otherwise a versus page.
The page is a fair comparison of who each option is for. Facts about the competitor come from its own website (Autopilot reads its homepage and pricing page) and public pages the writer checks, preferring the competitor’s own pages over third-party reviews. Prices and limits are marked “as of” the month the page was written. AI answers tell Autopilot what to beat, and it never takes facts from them. Every claim about prices, limits, plans and features is checked against its source. Unsupported claims are removed, and if one can’t be removed, the draft fails. The pull request has a Fact check section with how many claims were quoted from a source, and anything about your own product without a source, or any figure no source contains, is flagged for you to confirm. A flagged pull request never publishes on its own.
If your repository already has Markdown pages in a comparison folder (
compare, comparisons, content/compare, content/comparisons, content/vs or content/alternatives), the page goes there. Otherwise it goes in your blog folder.
Comparison pages don’t go through the planner, so they don’t wait for two signals, but they count toward the limit of three open pull requests. A closed comparison page can come back after 90 days, or sooner if, since you closed it, that competitor wins at least three more comparison questions and is among the three competitors costing you most.
If there’s nothing to write, the log says “No comparison page to write”, or “A comparison page is still waiting for review” while the last one is open.
Add internal links
Links your posts to pages ranking 4 to 20 in Google. A page just off the first page often needs only a few relevant internal links to move up.
Each links pull request comes together like this:
- It finds pages ranking 4 to 20 for queries with at least 10 impressions.
- It scans up to 60 of your posts, most seen first, for those exact phrases (two to eight words) where they aren’t linked yet.
- A review keeps only links that read naturally in their sentence. Wording never changes.
- It adds at most three new links per post and one per target page, checks every target against the live page, and opens one pull request with up to 10 files.
Update landing page copy
Rewrites landing pages AI cites less or Google ranks 4 to 20, changing only text.
Pages rendered from code (not Markdown, which the refresh and links automations handle) that lost AI citations by the same rule as refreshes come first, then pages ranking 4 to 20. It tries up to three and fetches each one live.
It changes only visible text: headings, paragraphs, button labels, list items, alt text and the page’s title and description. Code, links, class names and layout stay byte for byte, and a check rejects anything else. It makes at most 15 text changes and adds at most 250 words across up to five files. Copy in shared layouts only changes on the homepage. Facts come only from the live page, your own pages and the evidence, and it never claims to be the best or fastest. New copy with a figure no source contains, or a price or percentage only a competitor’s page states, is flagged and never publishes on its own.
After a landing page pull request, the next one waits 14 days so its effect can be measured.
A run without an update logs why, for example “No page lost AI citations or ranks 4 to 20 in Google”, “Every page losing citations or rankings has a Markdown source”, “The page’s copy isn’t in the repository” (for example when it comes from a CMS) or “None of the landing pages could be fetched”.
Related
Content gaps
How prompt and search gaps are found and ranked.
Measure results
See whether a post or update moved the number.