Bonus Level Studio

The AI Search Tracker Kit

Build a tracker that shows which Google searches and AI answers send people to your business, and what to do about it. Most of it runs on its own once it's set up. Free. No email required.

Get the kit on GitHubEvery file in one repo. If it helps you, star it. That is how other people find it.

Get each new part by email.

Why I built it

In October, Google showed Fuse in search results 27 times, down from 2,632 in May, and sent 2 clicks in the 90 days to November 1. Over August 1 to November 5, ChatGPT sent 305 people to Fuse and 17 signed up.

But when I asked ChatGPT the buyer's question by hand in September ("best YouTube analytics tools in 2026"), it named five other tools and not Fuse.

That gap is the whole point. AI assistants name the page that answers the exact question. If you only watch Google, you can't see which questions name you and which hand the buyer to someone else.

2,632to27
Google impressions
May to October
2
clicks
Google, 90 days to Nov 1
305
people
from ChatGPT, Aug 1 to Nov 5
17
signups
from those 305

The guide

What you need

  • Claude Code, which builds and runs the tracker.
  • Google Search Console, connected to Claude through a Search Console MCP server.
  • Your analytics, connected the same way. I use PostHog and its MCP server. Google Analytics works too.
  • A Claude API key and a Gemini API key, to ask the AI engines your buyer questions automatically. OpenAI and Perplexity keys are optional extras.
  • The two templates in this kit, if you'd rather see everything in a spreadsheet.

About two hours to set up. After that, the question runs and the reports happen on a schedule. The only hands-on part is a short monthly spot check.

Level 1: Connect your data (30 minutes)

Connect Search Console and your analytics to Claude Code through their MCP servers. Then Claude can pull the numbers itself instead of you exporting them.

From Search Console, the tracker pulls impressions and clicks by month, the pages that get seen but not clicked, and the search queries you already show up for.

From your analytics, it pulls the people who arrived from an AI assistant each month (chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, plus utm_source=chatgpt.com, which ChatGPT adds to many links), how many of them looked at a second page, how many signed up or became a lead, and where they landed.

It also runs a bot check. Real people use normal browsers and click on things.

Where AI visitors land tells you what they asked. If most of them land on one kind of page, that's the question the AI is answering with you.

Level 2: Build your buyer-question list (45 minutes)

If you write the questions yourself, you'll write the ones you already win. So the tracker starts from words buyers actually use:

  • Your Search Console queries. The searches you already appear for, pulled in Step 1.
  • Google, Bing and YouTube autocomplete, and Google's "related searches" and "People also ask". Free, and the tracker can collect them for you.
  • Perplexity's follow-up suggestions.
  • Reddit and forums, read by hand. Read, don't scrape.
  • Your own sales calls, emails and inbox. The best source. Add these yourself.

A caution on "AI prompt" keyword tools: at least some generate prompts from variations of your keyword instead of real typed questions. Use them for ideas, not for the list.

Claude turns the phrases into questions someone would ask an AI assistant, tagged by topic and by buyer stage: learning, comparing or choosing.

Then it draws the panel at random: 30 questions, 5 per topic, with a fixed seed it writes down. No swapping questions out afterwards. Make sure at least five are "choosing" questions, because that's where an AI names businesses.

Level 3: Let it run the questions (set up once)

On a schedule, the tracker asks every panel question through the Claude and Gemini APIs, and OpenAI and Perplexity if you add them. It asks each question several times, because answers change from run to run. For each answer it logs:

  1. Did it name your business?
  2. Did it link or cite your site?
  3. Which competitors did it name?
  4. Which websites did it cite?

It reports a share with counts ("named in 6 of 30"), never a single answer. If your name is a common word, it only counts answers that clearly mean you.

Once a month, spot-check by hand. Ask a handful of panel questions in the ChatGPT and Perplexity apps, logged out, and log them separately. API answers can differ from what people see in the chat apps, so the two are never mixed.

Level 4: Get your top three moves

The tracker reads your numbers and the run log side by side, and ranks the three changes most likely to matter. These are the patterns it looks for:

  • Google can't see your pages. Search Console shows pages discovered but not indexed, or your sitemap fails to load. A page Google can't see can't rank, and AI assistants that search the web can't find it either.
  • AI visitors arrive and leave. Give each landing page a next step that fits why they came.
  • You're absent on "choosing" questions. These answers lean on directories, review sites and lists. Earn outside mentions: G2, the directories the AI keeps citing, newsletters.
  • You're cited but not named. Put a plain, direct answer to the question near the top of the page.
  • One competitor wins the same questions every week. Read the pages the AI cites for them. That's your brief.

A typical top three: fix whatever stops Google from seeing your pages, give AI visitors a clear next step, and earn mentions on the sites the AI already cites.

Change one thing at a time, write down the date, and keep the same panel. That's how you'll know what worked.

Level 5: Read the weekly report

Each week the tracker tells you what changed, with counts, and whether your top three moves should change. That's the part you actually read.

The Claude Code prompts in this kit build all of the above. Paste them in order.

The Claude Code prompts

Paste these into Claude Code one at a time, in order, from an empty folder. Each one builds on the last. Replace anything in [BRACKETS].

Have these ready: your domain, access to Search Console and your analytics, a Claude API key and a Gemini API key. Keep keys in a .env file. Never paste a key into the chat. When Claude Code needs access, it will tell you what to set up yourself.

Prompt 1: Set up the project

I want to build an AI search tracker for my business, [BUSINESS NAME], website [DOMAIN].
We sell [WHAT YOU SELL] to [WHO BUYS IT].

Set up a small project in this folder:
- a data/ folder for CSV files and a runs/ folder for raw AI answers
- a .env.example listing the keys I need (ANTHROPIC_API_KEY, GEMINI_API_KEY, and optional
  OPENAI_API_KEY and PERPLEXITY_API_KEY), and .env in .gitignore
- a config file with my domain, business name, and the names people might use for us
  (include common misspellings, and note if our name is a common word)
- a README that says what each file is

Don't build anything else yet. Show me the plan for the next steps when you're done.

Prompt 2: Connect Search Console and my analytics

Help me connect Claude Code to Google Search Console and to my analytics ([PostHog / Google
Analytics / other]) through MCP servers. Tell me which MCP servers to use, and walk me through
the parts I have to do myself (installing, signing in, granting read access). Read-only access
is enough.

Then pull, and save to data/:
- gsc.csv: impressions and clicks by month for 6 months, the top 50 pages, and the top 100
  queries marked brand or non-brand using the names in my config
- ai_visitors.csv: people whose first visit came from an AI assistant, by month for 3 months
  (referrer or utm_source containing chatgpt.com, openai, perplexity, claude.ai,
  gemini.google.com, copilot), how many viewed 2 or more pages, how many signed up
  (our signup event is [EVENT NAME]), and their top 20 landing pages

Run a bot check on the AI visitors: what share used a normal browser, what share clicked
anything. Flag anything that looks automated, and tell me what you can't see in my data.

Prompt 3: Collect real buyer phrases

Build data/phrases.csv with columns: phrase, source, source_url, topic, buyer_stage
(learning, comparing, choosing).

Start with my own non-brand queries from data/gsc.csv. Then add phrases for these seed topics:
[3 TO 6 TOPICS YOUR BUYERS CARE ABOUT], from free sources only: Google autocomplete, Google
related searches, Bing autocomplete, YouTube autocomplete and Perplexity follow-up suggestions.
Keep the wording exactly as found. Do not invent phrases, and do not scrape Reddit.
Aim for 100 to 150 phrases. I'll add phrases from my sales calls and emails after,
marked source=customer.

Prompt 4: Draw the question panel

From data/phrases.csv, write data/questions.csv: one question per phrase, written the way
someone would ask an AI assistant. Keep the buyer's words.

Then draw a panel of 30 questions into data/panel.csv:
- 5 per topic (or as even as the topics allow)
- at least 5 "choosing" questions
- use a fixed random seed, write the seed in the README, and draw once
- no swapping questions out after the draw

Explain how the draw worked so I could repeat it.

Prompt 5: Run the panel through the AI engines

Write a script that asks every question in data/panel.csv through the Claude API and the
Gemini API (and the OpenAI and Perplexity APIs if those keys are set in .env). Ask each
question [3] times per engine, with web search turned on where the API supports it.

Save every raw answer to runs/<date>/. Then log one row per answer to data/run_log.csv with:
date, engine, question_id, question, run, named_us (yes/no), cited_our_site (yes/no),
competitors_named, sites_cited.

Count a mention only when the answer clearly means us, using the names in my config.
Show me the cost of one full run before you run it, and keep it under [$ LIMIT].

Prompt 6: Set up the monthly spot check

Set up data/spot_check.csv with the same columns as run_log.csv plus "app".

Give me a short monthly checklist: pick 8 panel questions at random, ask them by hand in the
ChatGPT and Perplexity apps, logged out or in a private window, and log what I see.

Keep spot checks separate from the API runs in every report. API answers can differ from
what people see in the apps, so never mix the two.

Prompt 7: Build the dashboard and my top three moves

Build a single-page dashboard (one HTML file I can open locally) from everything in data/:
- Google impressions and clicks by month
- AI assistant visitors, engaged visitors and signups by month, and where they land
- API results by engine, with counts (for example "named in 6 of 30"), and the spot checks
  shown separately
- the competitors and sites the AI names most

Then add "Top three moves". Rank the changes most likely to matter, using only my data.
Check for: pages Google can't see (not indexed, sitemap failing); AI visitors who leave without
a next step; being absent on "choosing" questions (needs outside mentions: G2, directories,
newsletters); being cited but not named; one competitor winning the same questions every week.

For each move, say what in my data points to it. Don't recommend anything my data doesn't support.

Prompt 8: Put it on a schedule

Make this run on its own every week: refresh the Search Console and analytics numbers, run the
panel, update the dashboard, and write a short report of what changed since last week, with
counts, and whether my top three moves should change.

Set it up as a scheduled job and tell me exactly how to check that it ran. Keep the weekly report
to one short paragraph plus the updated dashboard.

The templates

Two spreadsheets, if you'd rather see everything by hand. Open them in Google Sheets or Excel, delete the example rows and start filling them in.

Get the next part

This is part one of a series on building your own SEO and AI search tools. Get each new part by email.

Rather have it done for you?

I set this up for businesses: the numbers, the question panel, the automated runs and a short read on what to change. Email [email protected] and tell me what you sell.