How to Use Claude, Jev and Google Search Console to Find the Searches That Actually Pay You

Your Search Console chart went up and to the right all summer. Then it stopped. Nobody on the team can say why, and nobody can say whether the people who saw your site were ever going to buy.

That was us this month. So we ran the audit this guide teaches on our own site, startupmiracle.com, using three tools: Google Search Console for the data, Claude to do the work, and Jev, a new "decision" model from TypeSafe AI, to judge every single search query. The answer took an afternoon and it was not pretty.

• Google showed our site 6,425 times in 28 days. We got 41 clicks.
• 74% of those impressions came from developers, people who will never hire an AI consultant. They took 100% of our clicks.
• Searches from actual buyers, like "ai bdc for car dealerships" or "ai consultant Florida", made up about 21% of impressions and got zero clicks, because we ranked on page 3 to page 8 for them.
• One blog post carried 53% of all impressions. When interest in its topic cooled, the whole site cooled with it.

If you own a business, lead marketing, or run a performance team, this guide shows you how to run the same audit on your site, step by step, and what to fix first.

Why your SEO report is lying to you (a little)

Search Console tells you what people searched. It does not tell you who they were or whether they could buy. A report full of impressions feels like progress. But 1,000 impressions from the wrong audience is worth less than 20 from owners who are ready to pick up the phone.

The usual fix is a spreadsheet and an intern tagging queries by hand. That works for 50 queries. It breaks at 500, and it breaks again next month when the queries change.

It gets worse in the AI search era. Google AI Overviews, ChatGPT and Perplexity now answer many questions before anyone clicks (we covered the 42% click drop from AI Overviews earlier this year). Every click you still earn matters more, and every buyer search you rank badly for costs more.

What Is Jev? TypeSafe AI's Decision Model, Explained

Jev is a new kind of AI model from TypeSafe AI, released on September 15, 2026. Most AI tools you know, like ChatGPT or Claude, are built to write: they generate answers, emails and code. Jev is built to decide. You hand it a piece of text (a search query, a lead form, a support email) and a question with fixed answers, and it returns a typed answer your software can act on right away:

• Yes or no ("Is this lead urgent?") comes back as a probability, like 0.96.
• Pick one ("Which offer fits: Assessment, AI OS, or not a fit?") comes back as the winning option plus a probability for every option.
• A score ("How ready to buy is this person, from browsing to ready?") comes back as a number on the scale you define.

TypeSafe calls Jev a "System One" model, after the idea of fast, intuitive thinking. It was trained with a new method TypeSafe calls RLCD: reinforcement learning for calibrated decisions. Calibrated means the probabilities are meant to be honest. When Jev says 0.9, it should be right about 9 times out of 10. That is what lets your team set simple rules like "act automatically above 0.8, send everything else to a human."

Who built it? TypeSafe AI is led by co-founder and CEO Diogo Almeida. According to TypeSafe's team page, Diogo co-invented RLHF and InstructGPT, the methods that led to ChatGPT and GPT-4, and previously worked at Google Brain. In his launch post, he explains the question that started it: "why have superhuman chat models not led to AGI?"

Why it matters for marketing teams: the judgment calls that decide your pipeline (Is this search from a buyer? Is this lead hot? Does this page answer the question?) used to need a person or an expensive chat model. Jev makes them fast, cheap and consistent enough to run on every query, lead and review, not just a sample.

The three tools, in plain English

Google Search Console (GSC) is your source of truth. It is free, it comes from Google, and it shows every query where your site appeared, where it ranked, and whether anyone clicked.

Claude is the analyst that does the work. With Claude Code (Anthropic's AI agent that runs on your computer), you ask in plain language: "pull my last 28 days from Search Console and compare it to the month before." Claude writes and runs the steps, then explains what it found.

Jev is the judge. It answers yes/no, pick-one and score questions about each piece of text in well under a second. On our test it answered three questions about a sales lead in 0.38 seconds. That makes it perfect for the boring, high-volume judgment calls that break spreadsheets: "Is this search from a buyer or a developer?" asked 437 times, the same way every time.

Think of it this way: Claude is the analyst who does the work, and Jev is the calibrated judge that keeps it consistent.

Step 1: Get Jev access and your API key

1. Try it before you code. Go to the TypeSafe console and sign in. If your account is not active yet, request access from typesafe.ai first. Once you are in, open the Playground, paste any customer email as the "state", and add a yes/no question like "Does this message express urgency?" You will see a probability come back instantly.
2. Create an API key at console.typesafe.ai/keys.
3. Store it as an environment variable, never in your code or a shared doc:

export TYPESAFE_API_KEY="your-key-here"

4. Install an SDK if your team writes code (optional, the plain API works too):

pip install typesafe-sdk          # Python
npm install @typesafe-ai/sdk      # JavaScript / TypeScript (Node 20+)

One warning we learned the hard way: searching "Jev API" shows several lookalike sites that are not TypeSafe. Only use typesafe.ai, console.typesafe.ai and docs.typesafe.ai, and only send your key to api.typesafe.ai.

Step 2: Give Claude the TypeSafe skill

A "skill" is a package of instructions that teaches Claude how to use a tool correctly. TypeSafe publishes one, so Claude writes correct Jev requests instead of guessing.

In Claude Code, run these two commands in your terminal:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

Using Codex or another agent? Run npx skills add typesafe-ai/skills --skill typesafe-ai and pick your agent.

Then start with one of TypeSafe's own starter prompts:

Using the TypeSafe skill, run some experiments using the TypeSafe API key that I've exported to TYPESAFE_API_KEY. Propose changes based on the most promising results.

In Claude Code you can also call it directly with /typesafe:typesafe-ai.

Step 3: Connect Search Console to Claude

You have two options:

• No-code: in Search Console, open Performance, pick the last 28 days, and click Export. Hand the file to Claude.
• Connected: give Claude access to the Search Console API through a connector (we use Composio, which handles the Google login). Now Claude can pull any date range, page, or query on demand.

Ask Claude for three pulls to start:

1. Totals for the last 28 days vs. the 28 days before (clicks, impressions, CTR, position)
2. Every query with the page Google showed for it
3. A week-by-week trend for the last 3 months

One tip from Google's own Performance report guide: the last day or two is preliminary, and Google hides very rare queries for privacy. Your query table will not add up to your total. That is normal.

Step 4: Let Jev sort every search by intent

This is where the magic happens. For each query, Claude sends Jev the search phrase and two questions. The first picks one intent from a list you define. The second asks whether a business owner could plausibly be behind it.

Here is the exact question we used, trimmed:

{
  "state": { "search_query": "ai bdc for car dealerships" },
  "model": "jev-latest",
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": {
        "business": "Startup Miracle is a Miami AI consultancy for small and mid-size businesses...",
        "question": "Given `business`, what is the primary intent behind the search query?"
      },
      "criteria": {
        "buyer_services": "A business owner looking to hire or buy AI services",
        "local_buyer": "Looking for an AI company in a specific place",
        "developer_builder": "A developer researching tools to build with themselves",
        "irrelevant": "Unrelated to what we sell"
      }
    },
    "smb_buyer": {
      "type": "noul",
      "instructions": "Is the searcher plausibly a business owner who could become a paying client?"
    }
  }
}

Three rules that made the difference for us:

1. Describe your business inside the question. Our first draft without the business field was noticeably worse.
2. Always include an "irrelevant" or "none" option. Otherwise Jev must pick something, even when nothing fits.
3. Use confidence to decide what a human reviews. Jev returns how sure it is. We hand-checked the low-confidence ones and let the rest stand.

Our run classified 437 queries in a couple of minutes. Then Claude added up impressions and clicks by intent. That one table told us more than months of dashboards.

Step 5: Match each buyer search to the page Google shows

Next, for every buyer-intent search, look at which page is ranking. This is where the real problems show up. Ours:

• Searches like "ai startup consultants florida" landed on our About page, because we have no real local page. Google was improvising.
• "ai bdc for car dealerships" (77 impressions) landed on our dealership page at position 46. The page never uses the phrase "AI BDC".
• A flooring question about scheduling and customer communication ranked at position 6.5 with 50 impressions and 0 clicks. We were on page 1 and still not getting the click.

A search on the wrong page is a mapping problem, not a reason to spin up ten new URLs. Fix the page that should win first.

Step 6: Fix the cheap things, then log every change

Start with fixes that take minutes:

• Titles and descriptions. Our audit caught that every industry page title read "... | Startup Miracle | Startup Miracle". Duplicate branding wastes the words searchers actually read. Fixed the same day.
• Missing titles. Three of our solution pages showed the homepage title in Google. Also fixed.
• Answer the exact question. If buyers search "AI BDC", the page needs a section and a short FAQ that says "AI BDC" and answers it in the first two sentences.

Then keep an experiment log: date, page, one change, and the date you will re-check (use the same 28-day window). Change one thing per page, or you will never know what worked. Google's guide to impressions, position and clicks explains why one day of data is never enough.

Step 7: Track your AI mentions (the AEO part)

AEO means answer engine optimization: getting ChatGPT, Perplexity, Gemini and Google's AI answers to mention and recommend you. Two tricks:

Look for AI-shaped searches in Search Console. In our data, one query read: "small business ai consultant in miami. my location is miami, florida, united states, usa". No human types that. It looks like an AI assistant searching Google on someone's behalf, and we ranked at position 3.6 for it. When you spot queries like this, protect those pages.

Audit the answers themselves. Write down the 20 questions your buyers ask ("best AI consultant for car dealers in Florida"). Have Claude run them through ChatGPT and Perplexity each month and save the answers. Then ask Jev three questions about every answer: Is our company mentioned? Is it recommended? Is the description accurate? Now you have an AI-mentions score you can track like rankings.

TypeSafe cookbooks: recipes marketing teams can steal

TypeSafe publishes cookbooks, full worked examples with code you can hand to Claude. The How-to section covers search, formatting, tool selection and guardrails. Here is how each one maps to marketing work:

Two more sections worth a look: Batching, where TypeSafe shows that packing 13 questions into one request was 12.2x cheaper and 10x faster with the same answers, and Classification, which covers sorting content into deep category trees.

The fastest way in: point Claude at the cookbooks and ask, "Using the TypeSafe skill, which of these cookbooks fits a problem in my marketing stack?"

What results to expect

Be honest with your team: this audit does not guarantee rankings. What it does is replace guesswork with a clear list of which searches pay you, which pages should win them, and what to change first. We will re-run our numbers on October 25 with the same windows and publish what moved.

If you want the full playbook but not the setup, this is exactly what Startup Miracle does for small and mid-size businesses. Start with the free AI Readiness quiz. It takes a few minutes and shows where AI fits in your sales, marketing and operations. If search is your bottleneck, book the assessment and we will run this audit on your site.

Want the bigger picture of how we run our own company on AI agents? Read how we built our AI operating system with Claude Code and why Google now rewards pages that name exactly what they earn.

FAQ

What is Jev by TypeSafe AI?

Jev is TypeSafe AI's first "System One" model. Instead of writing text, it answers typed questions (yes/no, pick one option, or a score on a scale) with probabilities, fast enough to run on every query, lead or message your business handles.

Who created Jev?

Jev was built by TypeSafe AI, led by co-founder and CEO Diogo Almeida. Diogo co-invented RLHF and InstructGPT, the training methods behind ChatGPT and GPT-4, and previously worked at Google Brain. TypeSafe released Jev on September 15, 2026.

Is Jev a replacement for ChatGPT or Claude?

No. TypeSafe says so directly: Jev does not chat, write or code. Use Claude to do the work and Jev inside that work wherever a consistent judgment call is needed.

How do I connect Google Search Console to Claude?

Export your Performance report from Search Console and give Claude the file, or connect the Search Console API through a connector like Composio so Claude can pull any date range on demand.

What is the difference between SEO and AEO?

SEO gets your pages ranked in Google's links. AEO (answer engine optimization) gets your business mentioned and recommended inside AI answers from Google AI Overviews, ChatGPT, Perplexity and Gemini. You need both, and you can measure both.

How often should I run this audit?

Monthly, with the same 28-day windows, so you compare like with like. Log every change you make in between so you know what caused the movement.

Do I need a developer to do this?

Not for the basics. The Search Console export and the TypeSafe Playground need no code. Claude Code handles the scripting if you want the automated version. Startup Miracle can also set it up for you.