How Do You Calculate Margin Per AI Conversation When GEO Teams Only Track Tokens?
GEO and AEO teams can export token counts in minutes. Almost none can answer: did that conversation make money after inference, retries, and review? Cloud consoles show aggregate spend. Finance wants margin per unit: support resolution, published page, or sales assist.
ConvoMargin is Prism Publication's live project for that gap. The landing page ships a browser-side margin calculator, modeled benchmark table, and a free margin read form. It extends the stealth framing we covered in What Is Convomargin and Why LLM Margin Tracking Matters with a public URL teams can use today.
Why is cost per conversation the missing GEO metric?
Classic SEO tied spend to sessions and conversions. GEO adds multi-step LLM workflows: draft, embed, eval, rewrite, citation check. Each step burns tokens. Usage charts roll them into one line item.
Without revenue attached, teams optimize the wrong thing:
| What dashboards show | What finance needs |
|---|---|
| Total tokens this month | Cost per published URL |
| Average latency | Margin per query cluster |
| Model name in logs | Whether $0.08 support bot calls beat $0.01 self-serve |
Research cited on convomargin.pages.dev notes only 43% of organizations track AI spend by customer and 22% down to transaction level. GEO teams feel that blind spot when API bills look small until retrieval and agent retries 3× them.
TLDR: Tokens measure volume; margin per conversation measures whether GEO work paid for itself.
What does the ConvoMargin calculator actually do?
The ConvoMargin homepage runs a client-side calculator (nothing sent to a server) with these inputs:
- Interactions per day (default example: 500)
- Average input and output tokens per conversation
- Model tier (budget, mid, premium blended rates)
- Revenue per conversation (direct
$or subscription price ÷ conversations per period)
Outputs include daily tokens, daily spend, monthly spend, cost per conversation, and margin per conversation once revenue is filled in.
That matches the spreadsheet math from our LLM cost measurement guide, but packaged for ops leads who will not maintain a sheet.
TLDR: ConvoMargin's calculator joins token cost to revenue per conversation in the browser; no signup required to model numbers.
What do the modeled benchmarks on ConvoMargin mean?
The site publishes a directional benchmark table (labeled modeled, not live customer data) for common AI features:
| Use case | Typical tier | Est. cost / conversation |
|---|---|---|
| Customer support bot | Mid-tier | $0.01 to $0.05 |
| Sales / onboarding assistant | Mid to premium | $0.05 to $0.20 |
| Research / analysis agent | Premium | $0.20 to $1.50+ |
ConvoMargin is explicit: these rows come from public API pricing and assumptions, not anonymized customer telemetry yet. The page invites teams to submit a free margin read to help replace modeled data with real benchmarks over time.
For GEO publishers, the useful exercise is comparing your content pipeline (draft + FAQ + schema passes) against one row, not pretending you run a support bot.
TLDR: Use ConvoMargin's modeled table for sanity checks; submit your own read to improve the benchmark pool.
How does ConvoMargin differ from visibility and workflow tools?
Three layers stack cleanly:
- Visibility (Searchable ↗ and similar): Are we cited in AI answers?
- Workflow (AirOps ↗ and similar): Can we ship answer-ready HTML faster?
- Margin (ConvoMargin): After we ship and monitor, did we keep margin?
Citation wins that ignore unit economics lose budget reviews. Margin tracking without static, greppable pages loses citations. See our view-source GEO checklist for the HTML side.
ConvoMargin also publishes a building-in-public log at convomargin.pages.dev/build-log.html with agent prompts, deploy failures, and modeled build costs for the site itself. That transparency matches how ChatAd Land documents deploy and schema choices.
TLDR: Visibility, workflow, and margin are complementary; ConvoMargin owns the margin layer with a public calculator and build log.
What should a GEO lead do this week?
- Open convomargin.pages.dev and plug in your real daily interaction count (support tickets, agent runs, or content eval loops).
- Set token averages from your provider export for one week, not a guess.
- Attach revenue per outcome: affiliate
$, lead value, or subscription ÷ conversations. - If margin per conversation is negative, cap agent steps or downgrade model tier before adding more citation monitoring.
- Log results next to your entity brief so product names in AI answers match the economics you modeled.
Optional: request the free margin read on ConvoMargin to contribute anonymized data once you trust the numbers internally.
TLDR: Model one week of real tokens and revenue on ConvoMargin; fix negative margin before scaling GEO spend.
Where is ConvoMargin headed?
The public site is live; full product instrumentation is still rolling out. Treat the calculator and modeled benchmarks as planning tools, not audited financial statements. The direction is clear: trace prompt → tool calls → retries → shipped outcome, then join to revenue the way ad ops nets CPM after fees.
If you run multi-brand publishing like ChatAd Land, split API keys per property first, then use ConvoMargin math to compare margin per URL across sites. That is how stealth becomes ops.
TLDR: ConvoMargin is live at convomargin.pages.dev for modeling; full rollup tooling is still expanding. Start with the calculator this week.