1001 SEO Media
All posts
By 1001 SEO Mediaai rank trackingclient onboardingprompt trackinggeo agency

How to do AI rank tracking: our client setup, step by step

How we set up AI rank tracking for a new client: kickoff, building the prompt set together, competitors, engines and markets, the baseline, and monthly reporting.

"AI rank tracking" is a slightly wrong name for the job. AI answers do not have ranks the way a results page does. They name some brands, skip others, and cite a handful of sources. But it is the phrase clients search for, and the job behind it is real: find out whether AI engines recommend you for the questions your buyers ask, and whether that is changing.

This is the setup we run for every new client, in the order we run it. Each step says what we do with the client and what we configure in Promptwatch, the platform we run client programs on. If you are still choosing a tool, start with our ranked pick for 2026 or the agency comparison.

Step 1: Kickoff, and the decision the data has to support

Before anyone opens a dashboard, we ask the client what decision this data should help them make. Budget for content? Whether to enter a new market? Whether a competitor's push is working? The answer decides almost everything that follows: which prompts, which markets, who reads the report.

The rest of the kickoff is practical. Which products and segments matter this year. Which countries and languages. Who they lose deals to. What their brand is called in each market, including product names and legal suffixes. Who runs the website, which CDN sits in front of it, and who approves a tracking script. Our full list is in the onboarding questions we ask a new GEO client, so we will not repeat it here.

Step 2: Build the prompt set with the client

We build the prompt set together with the client, never for them. They know how buyers talk. We know how to turn that into a panel that measures something.

We start from their language: phrases from sales calls, support tickets, and the questions prospects ask in demos. Then we widen it with Promptwatch's Prompt Explorer, which suggests prompts from the client's Search Console data, their ranking keywords, People Also Ask, or any page on their site. Prompt Volumes adds monthly volumes, difficulty scores, and citation rates, which helps when the client has 200 candidate prompts and room for 50.

Every prompt gets a type: organic category questions, brand-specific questions, or competitor comparisons. We also tag intent, because a how-to prompt cites different sources from a "best X for Y" prompt. Prompts go into topics, which is how the client will read the results later. If the client sells to buyers who ask very different questions (a CFO and an IT lead, say), we add personas, and persona stacking when a buyer is both.

Size matters, and it is a budget question. On Promptwatch's brand plans, the response allowance works out to the prompt cap checked daily on about four engines: 50 prompts and 6,000 responses on Essential, 150 and 18,000 on Professional, 350 and 42,000 on Business. Agency plans have unlimited prompts with a shared response pool, so the math moves to the roster level. Either way, we size the panel first and add engines second.

When the client signs off, the panel is frozen for the month. New prompts get added at the monthly review, tagged so they can be filtered out of long-run charts. More on how we weigh volume against difficulty in how we build client prompt sets.

Step 3: Competitors and what counts as a mention

We add the competitors the client named, a few indirect ones, and then check Promptwatch's auto-discovered competitor list with the direct-competitor filter. That list can surface a rival the client never mentioned, which is a finding in itself.

Then we define the brand. Brand aliases catch spellings and product names. The Brand Book holds analysis instructions of up to 1,000 characters that decide what counts as a mention. Those instructions only apply to new responses, so we write them before the first run. Writing them in month two means month one was measured under different rules.

Step 4: Engines and markets

We pick engines by where the client's buyers are, not by how many a plan allows. Promptwatch monitors ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Microsoft Copilot, Amazon Alexa, and Meta AI through their real interfaces. Claude, DeepSeek, Mistral, and Grok are tracked through APIs, which shows what those models answer from training data. Both are useful, but they answer different questions, and the client report labels them separately.

Each market gets its own monitor. A monitor's language and location are fixed once created, which keeps a Dutch series from ever being blended with an English one. Country-level tracking is on every paid plan. State and city tracking is on Business and every agency plan. For local clients on Business, Agency Scale, or Enterprise, coordinate targeting pins a monitor to an exact point with a radius from 20 m to 5 km. ChatGPT uses the exact point, while Google AI Overviews and AI Mode use the nearest named area, which we explain to the client up front.

Prompts run daily by default. Business, Agency Scale, and Enterprise can run every 12, 6, or 1 hours. Daily is enough for nearly every client we take on, and faster runs use responses that could go to more prompts.

Step 5: Connect the data that explains the numbers

Visibility numbers tell you what happened. They do not tell you why. So in the first week we connect the layers that do, because they only start collecting once connected.

Crawler logs come first. Agent Analytics shows IP-verified visits from bots such as GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, and PerplexityBot, and a crawl-to-citation path with a citation rate per page. It connects through Cloudflare, AWS CloudFront, Fastly, Vercel, Netlify, Akamai, Google Cloud CDN, or the WordPress plugin. Crawler logs start on Professional for brands and are on every agency plan.

Visitor analytics comes second: a lightweight script, the GTM template, or the WordPress plugin, measuring AI-referred visits and conversions. ChatGPT stopped sending a referrer on 8 October 2026, so we switch on "Include ChatGPT UTM referrers", which counts utm_source=chatgpt.com visits as ChatGPT again. Google Search Console and Bing Webmaster Tools go in last, so AI data sits next to classic search data per page.

Step 6: The baseline is a month, not a day

One day of AI answers is a small sample. We treat the first full month as the baseline and do not report movement to the client until a second month exists to compare against.

During that month we mostly read. Which prompts never mention the client? Which pages get cited for competitors? We also look at query fan-outs per model, the sub-queries an engine ran before answering, because they show what content the engine went looking for.

We also keep notes on engine changes, because engines change without warning. Promptwatch's data report on ChatGPT's site: fanouts shows that on 8 August 2026, site: queries went from 0.37% to 16.8% of all ChatGPT fanout queries in a single day. A baseline month that spans a shift like that needs a dated note next to it, or the client will read the platform's change as their own progress or decline.

Step 7: The monthly rhythm

Between reports, one person does a short weekly pass on citations and crawler errors. We described that routine in how we track client citations in ChatGPT.

The monthly report reads the metrics together rather than one at a time. Visibility Score (0 to 100) is the headline. Share of Voice, Position, sentiment, citation rank, and self-citation rate explain it. Low Share of Voice is a coverage problem. A late Position is a prominence problem. Mentions without self-citations point to content or crawlability. Weak sentiment is a reputation problem. Each diagnosis leads to a different kind of work, so the report always says which one we are looking at.

We filter by prompt type before reading any trend, because adding comparison prompts can pull a blended number down without anything changing in the market. The Action Board gives us a prioritized list of what to do next, and Agent Chat answers the follow-up questions a client asks in the meeting. One detail for anyone comparing old reports: since 13 November 2026, Promptwatch compares a selected range with a previous period of the same length. Before that, "Last 14 days" was compared with 15 days. Our meeting format is in our monthly AI visibility review agenda.

Step 8: What we hand the client

At the end of setup, and every month after, the client gets things they can keep:

  • The prompt set as a document: every prompt with its type, intent, topic, persona, market, and the date it was approved.
  • A dated change log of everything we changed on their site, listings, and outreach.
  • The monthly PDF report, plus a dashboard or Looker Studio view for the people who want to dig.
  • A short action list: pages to refresh, prompt clusters with no page at all, pages crawlers fail to fetch, and third-party pages worth outreach.
  • An export of their data in CSV or Excel if they ever leave. If they hold the account themselves, the history simply stays with them.

Our recommendation

AI rank tracking works when the setup is boring and consistent: a signed-off prompt set, fixed monitors, a defined brand, crawler and visitor data connected from week one, and a baseline month before any claims. We do all of it in Promptwatch because the prompts, citations, crawler logs, and AI traffic sit in one project, so the monthly report can answer why and not only what. For our wider view of trackers, see what we use and why.

If you want a first look before any setup, run your domain through Promptwatch's free AI Visibility Report. It scans ChatGPT, Gemini, and Perplexity and checks crawler access, which gives you something concrete to bring to a kickoff. When you want us to run the setup above for your brand, contact us.