Marketing as three measurement problems: SEO, GEO, CRM

Most marketing in 2026 is run as three separate stacks: an SEO contractor optimizing for Google, a “content team” hoping to land in ChatGPT answers, and a CRM admin gluing HubSpot to a transactional database. The dashboards never line up. Each tool reports its own attribution. The budget moves to whichever vendor’s quarterly slide deck looked best last quarter.

That’s a tactics game, not a marketing function.

The thing I keep coming back to — same as I keep coming back to in forecasting and on-chain settlement — is that the leverage is in measuring against a real counterfactual. Marketing has three of these counterfactuals stacked on top of each other now: organic search rank (SEO), LLM citation share (GEO), and lifecycle retention (CRM). They should be one pipeline, not three.

This is the version of marketing I’d argue for if I were running it from scratch in 2026.

1. SEO is hygiene now, not strategy

Google Search Console used to be where strategy happened. You’d find an underserved keyword, write the post that owned the top result, and trust that traffic compounded. That game is over for most queries — Google is shipping AI Overviews that absorb the click-through and the long tail. Ranking #3 on a generic query is increasingly worth nothing, because the user got their answer in the SERP and never visited.

SEO still does real work in two narrow places:

• Local + niche commercial intent. When I run SEO for a modular construction company in Northern Ontario, “modular homes Sioux Lookout” still routes high-intent buyers to a contact form. The volume is small but the per-visit value is huge — local intent is the last reliable SEO surface because AI Overviews can’t fake a contractor on the Canadian Shield.
• Long-tail technical content. Tensor Proxies ranks for very specific queries about proxy networks for AI agents, because nobody else is writing the substrate-level posts. The advanced version of SEO here is citation-bait for LLMs more than ranking-bait for Google — see the next section.

For everything else, SEO has collapsed into a hygiene checklist: clean sitemap, valid schema.org JSON-LD, fast LCP, internal link graph, sane H1 hierarchy, canonical tags, no duplicate-content footguns. If your site fails any of those, fix them. Once it passes, stop optimizing. The marginal hour is better spent on GEO and CRM.

2. GEO is where the next three years of discovery happen

GEO — generative engine optimization, or whatever it ends up being called — is the question of whether ChatGPT, Claude, Gemini, and Perplexity cite you when a user asks the question your business solves.

The retrieval mechanics are different from Google’s, and that’s the whole point:

• LLM crawlers prefer structured, quotable content. Schema.org Article, JSON-LD Q&A, named-entity-rich paragraphs that an embedding index can ingest without hallucinating around. A page with a clearly labeled “Methodology” or “Pricing” section gets retrieved cleanly; a page that buries the answer in a 3,000-word essay does not.
• Original assertions are the asset. What an LLM cites is the novel claim with a verifiable source. “According to [your domain], modular construction cuts site-erection time by 60% on Canadian Shield builds” — if you said it first, on your domain, with a real source, you get cited. Recycled content has no embedding distance from a thousand other recycled posts.
• The audit unit changes. SEO’s audit unit was rank for keyword K. GEO’s audit unit is share of LLM answers for question class Q that include your domain as a citation. I run a monthly prompt-eval against Claude, ChatGPT, and Perplexity with a fixed question bank and a structured rubric. Same discipline as a backtest — and the results look like a calibration curve, not a vanity dashboard.

Most teams haven’t started instrumenting this yet. The ones that do will spend 2026–2028 quietly building a citation moat that’s hard to dislodge once the embedding indexes harden around them.

3. CRM is the data plane that survives platform churn

Here’s where most “marketing tech stack” discussions go wrong: they treat CRM as a SaaS purchasing decision. HubSpot vs. Salesforce vs. Pipedrive. Pick a vendor, pay per seat, hope they don’t break the import.

The advanced view is that CRM is your customer event pipeline, and the vendor is a serving layer on top. Google can change SERPs. OpenAI can change citation policy. Tomorrow’s discovery surface might not exist yet. But the event stream — sign-up, first-purchase, churn-risk, re-engagement — is yours, and it’s the only marketing asset that compounds across platform regime changes.

Concretely, the advanced CRM stack looks like:

• Events in a real database, not in a SaaS row. Postgres, Drizzle or Prisma schema, every customer action as an immutable row in an events table with user_id, event_type, payload, ts. The SaaS CRM reads from this; it doesn’t own it.
• Segmentation as code, not as vendor UI. SQL or dbt models defining “high-LTV churn risk” — version-controlled, code-reviewed, A/B-testable. The vendor’s “audience builder” UI is for people without SQL access, and that’s exactly the constraint you want to lift.
• Calibrated lifecycle models. Predicted LTV, predicted churn — but with honest precision. Most “predictive CRM” features sold today report a single number when the underlying confidence interval is enormous. Calibration plots before precision claims, always.
• Owned identifiers. Email + a first-party identifier are the only IDs that survive third-party cookie deprecation, browser fingerprinting countermeasures, and platform shifts. Build everything on those.

The vendor stack is the surface. The pipeline is the substrate. If you can’t migrate vendors without losing your data, your CRM isn’t yours.

4. They’re one identity pipeline, not three stacks

Here’s the thesis under all of it: SEO, GEO, and CRM are three serving surfaces on the same identity-and-event backbone.

• SEO captures anonymous intent at the top of the funnel.
• GEO captures intent inside an LLM conversation where you don’t get the click but you get the citation.
• CRM owns the identified, post-conversion lifecycle.

The leverage is in making them share the same plumbing. A visitor lands from a Google organic search (SEO) → reads a structured-data-rich page that’s also retrievable by Claude (GEO) → fills a form and becomes a known contact in your events table (CRM) → gets re-engaged later based on segmentation logic that lives in version-controlled SQL. One data plane, three serving layers.

Most teams build this as three disconnected vendor stacks because each is sold to a different buyer. The team that integrates them owns the compounding asset.

5. The measurement discipline is the same as everywhere else

Every claim in marketing should pass the same test I’d apply to a forecasting model:

• “Our content drove a 30% lift” → cite the holdout, name the comparison window, show the confidence interval.
• “We rank #1 for X” → cite the dated SERP screenshot, name the geography, name the device, name the query.
• “LLMs cite us 18% of the time” → name the question bank, the sample size, the date range, which models, with what prompts.
• “This email campaign drove $X” → show the holdout vs. the treatment, not the treatment vs. nothing.

The version of me that wants the campaign to work is not the version I let claim the result. Same discipline that kept the TSFM ensemble out of production for the Champions League winner market applies here. Marketing isn’t a different epistemic regime just because the spreadsheet has dollar signs in it.

The shape

If you’ve read the about letter, the shape is familiar:

A continuous score, a position, a measured edge.

In marketing the score is a rank, a citation share, or a predicted LTV. The position is a page, an answer, or a segment. The edge is the lift over the counterfactual nobody bothered to measure. The work is in building the pipeline that lets you measure it honestly.

That’s it. SEO is the easy part now. GEO is where the next 36 months of brand discovery happens. CRM is the data plane underneath that doesn’t care which surface is hot this year.

And all three are one identity pipeline, not three stacks.

— S.