Agentline

Who it was for

Software companies that want AI assistants to name them when buyers ask for tools like theirs.

What I did

• Started it after seeing, while working at Sperm Racing, that there wasn't much of a GEO presence
• Built the audit: it sends the questions buyers ask to the OpenAI, Gemini and Claude APIs, keeps every answer, and records who gets named and which sources get cited
• Built the diagnosis: for each answer that left a company out, Claude picks a reason from a fixed set (a competitor was chosen instead, the question was read as a different category, no strong candidate, not in its training data, or a factual gap). The product labels it as AI inference, not measurement
• Built Boost, the agent that drafts content from verified facts, publishes it to sites like WordPress and DEV.to, and checks it actually went live. It can also decide to do nothing: when the question is outside what the company offers (an AI judgment, labeled that way), or when the company was already named in every captured answer (measured)

Who did what

Me

Co-founder. I built the product (the audit, the diagnosis and Boost) and ran the business side.

Two co-founders

A fellow Haas student and an M.E.T. student.

AI coding agents

Wrote a big share of our commits, with us directing and reviewing them.

Challenges

• Startups in our target market (our ICP, or ideal customer profile) were often bringing GEO staff in-house instead of outsourcing it, and the ones we met that did outsource were already working with established providers like Profound

One question, start to finish

From the code: when Boost recommends nothing

Real excerpts from the Agentline codebase, trimmed where marked.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

export function decideBoostIntervention(
  input: BoostDecisionInput,
): BoostDecision {
  const { gap, brand, aiAssessment: ai } = input;
  const measured = gap.measured;
  // … build shared context and pick destinations …

  if (
    ai !== null &&
    ai.queryFit === "does_not_fit" &&
    !measured.questionNamesBrand
  ) {
    return noAction(
      shared,
      diagnosis(
        shared,
        "query_outside_offering",
        "ai_inferred",
        `${shared.observed} AI assessment (inference, not measured): ${ai.queryFitRationale ?? ai.rationale}`,
      ),
      `No content is recommended. The AI assessment judged this question to be outside what ${brand.name} offers, so content written for it would be off-message and unlikely to earn recommendations. If that judgement is wrong, reject this plan with a note and add verified facts or tracked competitors that show the fit before the next audit.`,
    );
  }

  if (gap.isFactualQuestion) {
    return decideFactual(shared, owned);
  }

  if (gap.outcome === "present") {
    return noAction(
      shared,
      diagnosis(shared, "already_visible", "measured", shared.observed),
      `No content is recommended: ${brand.name} was named in every captured response to this question.`,
    );
  }

  const decision = decideVisibilityGap(shared, owned, thirdPartyArticle);
  return withOpportunities(shared, decision);
}

What happened, when

1. 2025

   Working at Sperm Racing, I saw there wasn't much of a GEO presence, looked into the fixes out there, didn't like what I saw, and decided to build my own

2. 2026

   Built the audit, the diagnosis and Boost. Got 5 design partners and ran the agent 100+ times

3. 2026

   Early design partners said there was a big need. MyLife, a startup out of UC Berkeley, told us they were optimizing for GEO presence because that's where commerce is moving

4. 2026

   Met more startups in our target market: many were bringing GEO in-house, or already working with established providers like Profound

5. Oct 2026

   Audited Agentline's own brand: named in 0 of 89 completed AI answers, with 1 run failed

6. Oct 2026

   Wound down, because of what we heard from our target market

Evidence

Enlarge

Enlarge

Enlarge

Enlarge

Enlarge

Supporting context

Stack: Next.js and TypeScript on Supabase (Postgres), the Claude, OpenAI and Gemini APIs, and background workers · GEO (generative engine optimization): how often AI answers mention and recommend a company

What I learned