AI Search Tool Rank
All posts
By AI Search Tool Rank Teamai rank trackinghow tomeasurementai search visibility

How to do AI rank tracking: a measurement-first method in 7 steps

How to do AI rank tracking properly: define rank in an AI answer, build the prompt set, pick engines and cadence, read the metrics together and report monthly.

Most teams start AI rank tracking by typing their best keyword into ChatGPT and screenshotting the answer. That tells you what one engine said once, to one session, on one day. Run the same prompt an hour later and the list of brands can change.

AI rank tracking works when you treat it as a measurement problem first and a tooling problem second. Decide what you are counting, sample it often enough to average out the noise, and read several numbers together. The seven steps below follow that order. Where a step maps to a feature, we name the Promptwatch feature that does it, because that is the platform we'd run this method on.

Step 1: define what "rank" means in an AI answer

In Google, rank is one number per keyword. In an AI answer there are four things worth counting, and they behave differently.

Presence is whether your brand is named at all in a response. Position is where you land in the order of brands when you are named. Share of voice is your slice of all brand mentions across the prompt set. A visibility score combines presence and prominence into one figure.

The trap is reading position alone. Suppose a brand appears in 2 of 100 responses, both times first. Its average position is 1, which looks perfect, and its presence is 2%. Promptwatch's Position metric works this way too (it averages only across responses that mention you), so it is paired with a Visibility Score (0 to 100) built on different rules: each response is scored on placement, share of attention and depth, and responses without your brand count as zero. The Visibility Score can't flatter you the way position can.

Write your definition down before you buy anything: one response is one prompt, on one engine, in one run. Every metric below is an average over many of those.

Step 2: build a prompt set that looks like your buyers

The prompt set is your keyword list. Get it wrong and every chart after it is wrong too.

Start with the questions a buyer asks before they know your name: category questions ("best payroll software for a 50-person company"), problem questions, comparisons and alternatives. Add a smaller group of branded prompts so you can see how engines describe you. Tag each prompt by topic and intent, because you will want to report "comparison prompts" separately from "category prompts".

Keep it tight at first. A few dozen well-chosen prompts tracked properly beat several hundred you never read. Expand once you know which topics move.

This is where a tool saves real time. Promptwatch's Prompt Explorer suggests prompts from your Search Console data, your ranking keywords, People Also Ask, manual search, or any page on your site. Prompt Volumes adds monthly volumes and difficulty scores, so you can drop prompts nobody asks. Topics and tags let you group prompts for reporting, and personas let you run the same prompt as, say, a CFO and an IT lead. Persona stacking combines them.

Step 3: pick engines and countries

Track the engines your buyers use, not every engine that exists. For most B2B and consumer brands that means ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Copilot. Our guide to which LLMs to track goes deeper on the choice.

Then decide how each engine is checked. A live interface shows what a buyer sees, including the web results the engine pulls in. An API call shows what the model says from training data. Promptwatch monitors ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Copilot, Amazon Alexa and Meta AI through their real interfaces, and Claude, DeepSeek, Mistral and Grok through APIs. Keep those groups apart in your reporting.

Location matters more than people expect, because AI answers change by market. In Promptwatch, a monitor is a prompt panel with fixed models, country and language; its language and location can't be changed after you create it, so plan one monitor per market. Country tracking is on every paid plan. State and city tracking, and coordinate targeting within a 20 m to 5 km radius, start on Business, which is what multi-location brands need.

Step 4: set the cadence and do the response math

AI answers vary, so a single run per week leaves you guessing whether a change is real. Daily runs are the sensible default. Faster runs make sense for launches, crises or ad-heavy categories. Promptwatch runs prompts daily, and Business adds every 12, 6 or 1 hours.

Budget in responses, not prompts. Responses equal prompts times engines times runs. Fifty prompts on four engines, run daily for 30 days, is 6,000 responses, which matches the allowance on Promptwatch's Essential plan ($95 a month). Add engines or move to hourly runs and the number multiplies fast. Professional ($245) includes 18,000 responses and Business ($579) 42,000.

Step 5: read Visibility Score, Share of Voice and Position together

Each metric on its own misleads. Read together, they point at a specific problem. This is the reading we use, which follows Promptwatch's own guidance:

What you seeWhat it usually meansWhere to look next
Low Share of VoiceCoverage problem: you're absent from too many answersPrompts where competitors appear and you don't
Mentioned, but late PositionProminence problem: engines know you but rank others firstSources cited for the brands above you
Mentions without self-citationsContent or crawlability problemYour own pages and crawler logs
Weak sentimentReputation problemClaims engines make about you

Watch four-week trends, not single days. Promptwatch's A/B period comparison and heatmap (brands, models, topics or countries on either axis, colored by visibility, sentiment, position or share of voice) make the pattern visible quickly. Brand Narrative goes one level further and pulls the individual claims engines make about you and your competitors into themes such as pricing or support.

Step 6: diagnose with citations and crawler logs

Metrics tell you that something is wrong. Citations and logs tell you why.

Start with citations. For the prompts where you lose, look at which pages and domains the engine cited. Often it is a comparison page you don't have, a review directory, or a Reddit thread. Promptwatch's citation analytics show pages, domains and URL positions, with Reddit, YouTube, LinkedIn and Instagram reports, and query fan-outs show the exact sub-queries each model ran before it answered. Those sub-queries shift. Promptwatch's data shows that on August 8, 2026, site: queries went from 0.37% to 16.8% of all ChatGPT fanout queries (report). Keyword mappings built before a shift like that go stale.

Then check access. If your page is the right answer and still isn't cited, find out whether AI bots fetch it at all. Promptwatch's Agent Analytics logs IP-verified crawlers such as GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot, and shows a citation rate per page. A page that bots never fetch has an access problem: robots.txt, errors, or a CDN rule. A page that gets fetched and not cited has a content problem. Those need different fixes, and without logs you are guessing which one you have. Crawler logs start on the Professional plan.

Step 7: report monthly, and tie it to traffic

A monthly report should answer three questions: did visibility move, why, and did it reach the site.

Report the Visibility Score, Share of Voice and Position as four-week averages by topic and engine, next to your top three competitors. Add the citation changes that explain the moves. Then add the business line: AI-referred visits and conversions. Promptwatch's visitor analytics measure those on every plan, and since ChatGPT stopped sending a referrer on October 8, 2026, a toggle counts utm_source=chatgpt.com visits as ChatGPT again.

To keep the reporting from eating a day each month, use scheduled PDF reports from a custom dashboard (every brand plan includes 10), or the Looker Studio connector from Professional. The Action Board turns the findings into a prioritized to-do list, so the report ends with next actions instead of a chart.

Common mistakes

Tracking too many prompts too early is the most common. The second is mixing markets in one monitor, which averages a strong country with a weak one into a number that describes neither. The third is reporting position without presence. And the fourth is treating a one-day drop as news. Give every change two weeks before you act on it, unless your crawler logs show bots hitting errors, which you should fix the same day.

Where to go from here

If you want the tools ranked against this method, our 2026 AI rank tracker scorecard weights these same capabilities, and AI rank tracker compared shows which tools publish which features. The general ranking is at what are the best AI rank trackers, and our guide on how to measure your brand's AI search visibility covers the wider metric set.

Every step above maps to a feature in Promptwatch, from Prompt Explorer in step 2 to Agent Analytics in step 6. The quickest way to try the method end to end is the 7-day free trial on Professional, which includes crawler logs. Load your first prompt set on day one and you'll have a week of daily data before the trial ends.