How to do AI rank tracking: a loop that turns drops into fixes
How to do AI rank tracking that triggers fixes: build the prompt set, define what counts as a drop, route it to a task queue, publish, then verify with crawl data.
Most AI rank tracking setups fail quietly. The prompts run, the dashboard fills up, and once a month someone notices that a few numbers went down. Nothing happens next, because nobody decided in advance what a drop is, who owns it, or what the fix should be.
The method below treats tracking as the first step of a loop rather than a report. Prompts produce a signal, a threshold turns the signal into a task, the task becomes a fix (some written by an agent, some by a person), and the fix gets checked against crawl, citation, and traffic data. We use Promptwatch to show each step, because it covers the whole loop in one project, but the logic holds whatever you run it on.
Step 1: build a prompt set worth defending
Start with the questions you would be upset to lose. Pull them from three places: what customers ask sales and support, what your Search Console queries look like when written as full questions, and which comparisons buyers make between you and named competitors. Write each one as a person would type it into ChatGPT, not as a keyword.
Label everything. Promptwatch sorts prompts into types (Organic, Brand Specific, Competitor Comparison) and intents (from Informational to Transactional), and lets you group them into topics with tags. Those labels matter later, because a drop on a commercial comparison prompt deserves a faster response than a drop on a definitional one. Prompt Explorer suggests prompts from Search Console, your ranking keywords, People Also Ask, or any page on your site, and Prompt Volumes shows monthly volume, difficulty, and citation rates so you can start with prompts that are worth winning.
Decide on markets before you create anything. A monitor's language and location are fixed once it exists, so set up a separate monitor per country and language from day one. Country tracking is on every paid plan. State and city tracking start on Business.
Step 2: pick a cadence and respect the noise
AI answers change from run to run even when nothing about your site changed. A single bad day means very little. Daily runs are the sensible default, and Promptwatch runs every monitor daily, with every 12, 6, or 1 hours available on Business, Agency Scale, and Enterprise. Faster cadence costs responses, so save it for the prompts where hours matter, such as launches or news.
Read results over windows, not single runs. Seven days against the previous seven days is a good starting unit. It's also worth looking at query fan-outs, the background searches a model runs before answering. Promptwatch shows the exact sub-queries per model, and they often explain a drop better than the prompt itself, because the cited sources come from those searches.
Step 3: define what counts as a drop
Write your thresholds down before the data tempts you to rationalise. These are the rules we suggest as a starting point. They are our editorial defaults, not settings Promptwatch ships with, so tune the numbers to your volume.
| Signal | Suggested trigger | Likely problem |
|---|---|---|
| Visibility Score by topic | Falls by a set number of points, 7-day average vs previous 7 days | Coverage or prominence |
| Share of Voice | A named competitor overtakes you on a topic | Coverage |
| Position | Average slips on commercial prompts (lower is better) | Prominence |
| Page citations | A tracked URL stops being cited on prompts it used to win | Content or crawlability |
| Sentiment | Drops below the neutral midpoint (about 50) | Reputation |
The "likely problem" column follows how Promptwatch reads its own metrics together. A thin Share of Voice means you're missing from too many answers. A late Position means you're there but buried. Mentions without self-citations point to content or crawlability. Weak sentiment is a reputation issue, and no blog post fixes that on its own.
Use the Trends view (risers and decliners) and A/B period comparison to test a rule before you trust it.
Step 4: route drops to one queue
In Promptwatch that queue is the Action Board, a prioritized to-do list covering content gaps, sentiment, Reddit, offsite mentions, site health, and custom actions. A Discovery agent investigates on a weekly schedule, one focus per day, and suggests up to five evidence-backed actions per run. You can assign items, comment, and @-mention the agent for a reply, and a digest email arrives on Mondays.
Your own thresholds can feed the same board. A Scheduled Chat in Agent Chat can run the comparison every week, flag topics that crossed your trigger, and deliver the result in the app or in Slack. Agent Chat works the Action Board directly and asks for any missing value before it changes something. If you prefer code, the MCP server has read and write tools your own agent can call, including for Action Board items. Keep the active queue short. Five items with owners beat forty without.
For a concrete version of this routine focused on ChatGPT, see our weekly citation workflow an agent can run.
Step 5: decide which fixes an agent can do
This is where teams either save a lot of time or create a mess. Our rule: agents handle reversible, evidence-backed edits to content you own. People own anything that changes a claim, site access, or a relationship.
| Fix | Agent | Person |
|---|---|---|
| Refresh a page that lost citations | Drafts the update | Approves it |
| New article for a fan-out or content gap | Writes it | Reviews before publish |
| Page that states something AI repeats wrongly | Rewrites flagged pages | Confirms what is true |
| Internal links on new content | Picks and adds them | Spot checks |
| Crawler blocked or erroring | Flags it | Fixes robots.txt or server |
| Reddit threads, PR, partner pages | Finds the opportunity | Does the outreach |
| Pricing and positioning pages | Suggests | Writes and signs off |
Promptwatch maps onto the left column fairly directly. Content Refresh proposes diff-style updates for existing pages. Content creation can start from a tracked prompt, a query fan-out, or a cited page, using a brief that includes competitor URLs and your internal links, and it picks internal links before writing. On Business, Fact Check compares claims AI makes about you with your site and the most-cited offsite pages, and "Fix with content agent" refreshes each flagged page on your site, one article per page. The right column stays human. Crawler access problems, in particular, show up in Agent Analytics with failure explanations, but someone with server access has to fix them.
Step 6: publish through the Content Agent or your CMS
On Business, the Content Agent plans, writes, and publishes against visibility gaps on a calendar, with a monthly budget and publish windows. Pieces land in a review inbox where you approve, edit, or decline, or you can switch to fully automated mode once you trust the output. We would keep review on for at least the first month. On Professional you get AEO articles on demand (5 included) and publish them yourself.
Publishing connects to WordPress through the official plugin, to Webflow through OAuth, and to Framer with a project URL and API key, as a draft or live. Publish status tracking then follows each document from draft to live URL and on through crawls, citations, and AI traffic. If your site supports it, set up an IndexNow key in project settings so new and changed pages can be submitted from the Site view.
Step 7: verify the fix before you call it done
Set a review date when the fix ships, then check three things in order.
First, did AI crawlers fetch the page? Agent Analytics, from Professional, shows IP-verified hits from GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and others, with a citation rate per page. If the page was never crawled, you have an access problem and the content is irrelevant for now.
Second, is it being cited? Page Tracker and the AI Visibility tab on each sitemap page show one URL's citation rate and position over time, with the prompts and responses that cite it. If crawlers came but citations didn't, it's a content problem, and the page goes back through Content Refresh.
Third, did anyone arrive? Visitor analytics, on every plan, count AI-referred visits and conversions. ChatGPT stopped sending a referrer on 8 October 2026, so switch on the "Include ChatGPT UTM referrers" toggle to count utm_source=chatgpt.com visits again. Search Insights adds Search Console and Bing data per page for the classic search side.
Freshness affects how soon you should look again. Promptwatch's content half-life report found that half of the news pages ChatGPT cited on July 6 were no longer cited after 36 days, while 78% of landing pages and 76% of docs were still cited at 30 days. News-style pages need a shorter re-check cycle than evergreen ones.
Putting the loop on one platform
You can assemble this loop from a tracker, a log tool, a content autopilot, and an analytics script. Our comparison of monitor-only trackers and agentic platforms explains where those stacks leak: the writing tool never learns which prompts dropped, and nothing checks whether the result got crawled. Our ranked best AI rank tracker for 2026 and the broader list of the best AI rank trackers cover the alternatives.
Promptwatch is the practical choice because every step above happens in the same project: prompts and thresholds, the Action Board, the Content Agent, then crawler logs, Page Tracker, and visitor analytics to confirm the result. Start the 7-day free trial on Business ($579/mo) to run the full loop including automated publishing, or on Professional ($245/mo) if your team will write the fixes and you mainly need crawler logs and tracking.