Skip to content
Primetra

Method

What you get, and why it's hard.

Asking ChatGPT about your category takes a minute. A number you can act on, every week, takes a measurement system. Primetra runs that system and hands you the result.

Current method: found-v2 · used-v4 · fix-v1

Every week

What you get each week.

Three answers kept apart: what assistants say about you, what agents do on your product, and what to fix first. No blended score.

  1. 01

    Share of Agent across five assistants

    How often Claude, ChatGPT, Gemini, Grok and Perplexity cite you when your buyers ask, and which competitors they cite instead.

    Sample data

    Who gets cited

  2. 02

    The agents that showed up

    Which real agents found you, connected, called your tools, got data back and came back for more. Fakes and noise removed.

    Sample data
  3. 03

    Three ranked fixes, patches ready

    The three changes most likely to move your numbers, with the evidence, and a ready patch for your robots.txt, MCP tools or pages.

    Sample data
    /robots.txtPatch ready

    Patch for /robots.txt: remove “Disallow: /”, add “Allow: /”.

    Fix 1 Unblock OAI-SearchBot · re-check by Oct 26

  4. 04

    A verdict on every fix

    Mark a fix done and we re-check it when the numbers can have moved. Each one ends as verified, not yet, too few to read, or didn't work.

    Sample data

    Last week's fixes, re-checked

    • Unblock OAI-SearchBotVerifiedIt worked: Share of OpenAI bot requests you blocked went from 16% to 0% (n=43), a real change, not week-to-week wobble.
    • Block fake PerplexityBot trafficDidn't workIt went the wrong way: Requests pretending to be an AI bot moved from 14 to 26.
  5. 05

    A go-to-market plan

    Four weeks of moves built from real answers: the questions to own, the sources to earn, the agent channels to open.

    Sample data
    1. W13 tasks
      Open the doors
    2. W23 tasks
      Answer the questions
    3. W33 tasks
      Earn the mentions
    4. W44 tasks
      Measure and repeat

    Week 1 starts with: Unblock OAI-SearchBot.

    See the sample plan
06

A five-minute memo

All of it in one short email every Monday, with the full page behind it when you want the detail.

Open the sample report

Monday memo · Relaykit

From Primetra · Mon, Oct 19 · 5 min read

Sample data
Share of Agent

21%95% 17–24% · n=600

Biggest leak
Courierly is cited in 56% of 600 answers; you in 21%.
Ship this

Unblock OAI-SearchBotPatch ready

re-check by Oct 26

Also this week

  1. Fix 2: Make github.com/relaykit/relaykit-node worth citing

    An agent can draft it · a few hours · re-check by Nov 16

  2. Fix 3: Make relaykit.example worth citing

    An agent can draft it · a few hours · re-check by Nov 16

Last week

Fixed, and it worked.

Agent errors on send_email dropped from 14 to 0 (n=58 calls).

The hidden work

Why it's hard to do yourself.

You could ask ChatGPT about your category tonight. You'd get one answer, from one run, on one day, and no way to tell whether it holds next week.

  1. 01

    Scale and repetition

    30 buyer questions, five assistants, each question asked several times a week. That's hundreds of answers, because one run swings a lot. API answers also differ from what the apps show, so you have to know which one you're measuring.

  2. 02

    Real agents, not noise

    User agents lie. Each vendor publishes its IP ranges and signatures its own way, and changes them often. New agents keep appearing. And analytics only sees people: without a collector on your stack, agents are invisible.

  3. 03

    Statistics that hold up

    A rate needs a range, a minimum sample before anyone reads it, a noise band before a change is called, and a version so a rule change can't fake a trend. Skip one and you chase noise.

  4. 04

    From finding to fix

    Knowing a bot is blocked is half the job. The other half is a patch for your actual robots.txt, MCP tools and pages, then a re-check at the right time: days for bot and MCP fixes, a few weeks for citations.

  5. 05

    Upkeep, every week

    Bot tables go stale. Assistant APIs change models and search settings. New agents arrive. Someone has to notice, update the pipeline and keep last month's numbers comparable.

  6. 06

    Patterns across every site

    We maintain the noise rules and bot tables across every site we measure. When we add a rule for a new sweep or spoofer, every customer gets it, and the method version records the change.

Why we ask hundreds of timesSample data

n = 600 answers · 123 cite you · 20.5%

95% 17.5–23.9% · ±3.2 pts

Sample data. The rate stays near 21%; only the number of answers changes. With 20 answers the true value could sit anywhere from 8% to 42%. With 600 it's pinned to 17%–24%, close enough to act on.
Removed before we count
  • Discovery sweepsScripts that hit well-known paths across thousands of sites.
  • ProbesCalls with test arguments, sent to see what a tool does.
  • EnumerationCallers that list every tool, then leave.
  • Bursts and bulkOne caller hammering you far harder than real use needs.
  • Spoofed botsRequests that use a known bot's name without being that bot.
  • Your own trafficYour monitoring and health checks.

We count what we removed, by reason, so you can check our work.

Do it yourself vs Primetra

  • Asking
    Do it yourself

    Script hundreds of API calls a week across five assistants, repeat each question, and store every answer.

    With Primetra

    The panel runs every week. You read the result.

  • Seeing agents
    Do it yourself

    Build a collector for your edge, keep vendor IP ranges and signatures current, and filter sweeps, probes and spoofers.

    With Primetra

    One small collector, about five minutes to install. We keep the bot tables and noise rules current.

  • Reading numbers
    Do it yourself

    Work out ranges, minimum samples and a noise band, and version your rules so trends stay honest.

    With Primetra

    Every number arrives with n and a 95% range. Noise stays out of your inbox.

  • Fixing
    Do it yourself

    Turn each finding into a patch for your robots.txt, MCP tools and pages, then remember to re-check it.

    With Primetra

    Three ranked fixes a week, patches drafted, each one re-checked for you.

  • Keeping it running
    Do it yourself

    Track API changes, new agents and stale bot tables. Every week.

    With Primetra

    We do the upkeep and mark method changes on your charts.

Doing it yourself is a small engineering project, then a weekly job forever.

The trust layer

Our promises.

The rules every number follows, in your memo, on your page and in your exports.

  • Every number shows its sample size and a 95% range.n is how many answers or calls it rests on. The range is where the true value very likely sits. With too few samples you see “too few to read”, not a guess.
  • Empty is never shown as zero.Missing data is a dash with its reason, like “needs the MCP SDK”. A timed-out answer drops out of the count instead of counting as “not cited”.
  • A change is only called when it clears the noise.A move from 38% to 42% is usually noise. If a move stays inside the noise band, it stays out of your inbox.
  • Verified and inferred are labelled.A verified finding rests on direct evidence, like an error your MCP server returned. An inferred one rests on a pattern, and says so.
  • Official APIs only.We ask each assistant through its official API and never scrape the consumer apps. The ChatGPT or Claude app can answer differently: it may use memory, location, another model or other search settings. The APIs are repeatable and allowed.
  • Method versions are marked on charts.Every number carries the version of the method that produced it. When a rule changes, your chart shows the break instead of drawing a trend across it.
  • Ordinary visitors never leave your stack.The collector drops ordinary browsing before anything is sent. We see agents, bots and visits that come from an AI assistant's answer.

Customers also get a method appendix next to their own numbers: every threshold, window and definition behind them, so they can audit each one.

Sample data
Claude answers that cite you
39%
n=96 · 95% 29–49%
Grok answers that cite you
—
Too few to read · n=9 so far
Agents that got data back
—
No data · needs the MCP SDK

Sample data. Enough answers: a number, its n and its range. Too few: no number. No data: a dash and the reason.

Every bot gets one of three labels
  • VerifiedIt is who it says it is.
  • SpoofedIt uses a name it can't back up. We remove it and count it.
  • UnverifiableThere's nothing published to check it against. We keep it, label it, and never guess.

What we don't claim.

  • llms.txt is hygiene, not a ranking factor. Published tests (Ahrefs, SE Ranking, Otterly) found no measured citation lift, so we check for it and don't sell it as a fix.
  • We don't promise placement in any assistant. We show where you stand and what moved.
  • A free check is directional: 5 questions, asked once. The report and Monitoring ask every question several times.
  • Agent pick tests (which tool a coding agent chooses in a lab run) aren't part of v1.

See it on your own product.

The free check takes about two minutes. The sample report shows a full week.

Run the free check See a sample report