How Siteup Built an Attribution-Native GEO Program That Traces AI Citations to Closed Revenue

In early 2026, Siteup faced the same problem as every other GEO SaaS company: we were building software that helped customers get cited in AI answers, but we couldn't prove our own GEO program was generating pipeline. Our citation dashboard showed growing visibility. Our CRM showed nothing.

Twelve months later, we can trace AI citations to specific deals in our pipeline — not through a custom integration or an agency engagement, but by using the same agent automation architecture we ship to customers. Every piece of content we publish carries embedded metadata: funnel stage, keyword cluster, conversion endpoint, and freshness timestamp. When a deal closes and the buyer says "I found you through ChatGPT," our CRM traces the citation back to the specific article that earned it. Attribution is a report, not an investigation.

Here's how we built it, what broke along the way, and what we'd do differently.

The Challenge: A GEO Company That Couldn't Prove GEO Worked

Siteup sells agent automation software for GEO. Our value proposition is that agent-generated content with embedded metadata makes AI visibility attributable to revenue — unlike manual content programs where attribution requires post-hoc stitching across disconnected tools.

In January 2026, we were not using our own product for our own GEO content. Our blog was manually produced — writers created articles, an SEO manager tracked citations in a third-party monitoring tool, and nobody connected either system to our CRM. When potential customers asked "does your own GEO program generate pipeline?", we had citation counts and visibility scores. We didn't have a revenue answer.

The numbers from that period tell the story. Our GEO monitoring tool showed our brand appearing in roughly 200 AI answers per month across ChatGPT, Perplexity, and Gemini. Our share of AI voice for "GEO software" and adjacent terms was trending upward. Our Google Analytics showed approximately 0.4 percent of site sessions arriving from AI platforms — consistent with the industry pattern where analytics-tracked AI referral traffic dramatically undercounts actual AI influence.

And our CRM showed zero AI-attributed pipeline. Not because AI wasn't influencing buyers — we knew from sales calls that prospects were finding us through ChatGPT and Perplexity. Because we had never instrumented the connection. Our content system, our citation tracker, and our CRM were three separate tools operated by three separate people. By the time anyone asked "did that article generate pipeline?", the attribution trail was cold.

What We Tried First — and What Failed

Our first instinct was the obvious one: buy a better GEO monitoring tool. We evaluated tools with larger prompt databases, more frequent refresh cycles, and fancier dashboards. The demos were impressive. The pricing was reasonable — $295 to $500 per month for enterprise-tier monitoring. We almost bought one.

What stopped us was a simple question from our head of revenue: "Will this tool tell me which deals came from AI citations, or will it just count citations more accurately?"

The answer was the latter. Better monitoring would tell us we appeared in 342 AI answers instead of "roughly 200." It would show sentiment analysis and competitor share-of-voice comparisons. It would not connect any of those citations to the deals in our HubSpot pipeline. We would be paying more for a more precise measurement of a signal we couldn't trace to revenue.

The realization forced a different question: what's the minimum system that would actually answer the revenue question? The answer turned out to be a CRM field, a form field, and an SDR script — total cost $0 — plus a content architecture change that our own product already supported.

The Approach: Building Attribution-Native GEO From the Ground Up

Phase 1: CRM Instrumentation (Month 1)

Before touching our content production, we instrumented our CRM. Three changes, implemented in a single afternoon:

AI Discovery Source field. We added "AI Search," "ChatGPT," "Perplexity," "Claude," and "Gemini" as values in our HubSpot Lead Source property. Any lead where the buyer mentioned discovering Siteup through an AI tool was tagged accordingly.

Free-text attribution on signup. We added a "How did you hear about us?" field to our signup flow — free text, not a dropdown. Within two weeks, we started seeing responses like "ChatGPT recommended Siteup for GEO content automation" and "Perplexity comparison of GEO tools." These responses confirmed what our analytics couldn't see: AI was influencing our pipeline.

SDR discovery script. We added one question to our discovery call script: "Were you using any AI tools — ChatGPT, Perplexity, Claude — when you started researching GEO solutions?" Our SDRs logged the answers in HubSpot. Within 30 days, we had confirmed AI touchpoints on roughly 20 percent of our qualified opportunities — deals where the buyer explicitly said they discovered Siteup through an AI recommendation.

The CRM instrumentation took one afternoon and cost $0. It surfaced more attribution signal in 30 days than our GEO monitoring dashboard had produced in six months.

Phase 2: Switching to Agent-Automated Content (Month 2–3)

This was the harder phase. We moved our blog from manual production to our own agent automation pipeline. The key change wasn't the content quality — our agents produced content comparable to what our writers had been producing. The key change was the metadata layer.

Every article generated by the agent pipeline carried four embedded metadata fields:

1. Funnel stage — TOFU, MOFU, or BOFU. The agent determined this from the content brief's target audience and keyword intent.

2. Keyword cluster — the specific commercial-intent query cluster the article was designed to target. This created a persistent link between content pieces and the keyword strategies they served.

3. Conversion endpoint — the page the article was designed to drive conversions toward, typically our signup flow or pricing page.

4. Freshness timestamp — last publication and modification dates, automatically updated when the agent refreshed content.

This metadata layer is what made attribution possible. When a deal closed and the buyer's AI Discovery Source was "ChatGPT," we could query which keyword clusters were generating citations at the time the buyer would have been researching. We could trace from the keyword cluster to the specific content piece designed to earn those citations. We could verify whether that content was fresh at the time of the buyer's research. The full chain — content → citation → preference → branded search → conversion → CRM — was instrumented.

Phase 3: The Refresh Flywheel (Month 4–6)

The Stacker/Scrunch research finding that AI citations have a 4.5-week average half-life meant our content needed continuous refresh. Our agents were scheduled to review and update each article on a rolling basis — updating dateModified timestamps, refreshing statistics, and restructuring content for continuing citability based on the GEO Measurement Study's findings (explicit dating: +22 percent citation lift; chunk-level structure: +18 percent).

The refresh flywheel worked as follows: agents refreshed content → fresh content earned new citations → new citations generated attribution data → attribution data informed which keyword clusters and funnel stages drove the most pipeline → the content brief model prioritized those clusters for additional content and more frequent refresh.

By month six, we had a growing dataset: which content pieces, targeting which keyword clusters, at which funnel stages, drove which conversion endpoints, resulting in which pipeline. Attribution was a query, not a forensic investigation.

Results

Our results reflect the architectural shift from post-hoc to attribution-native — not a sudden explosion in traffic or citations, but a transformation in what we could measure and act on.

Attribution infrastructure: Before the switch, our CRM showed zero AI-attributed pipeline. After CRM instrumentation (Month 1), we identified AI touchpoints on approximately 20 percent of qualified opportunities — roughly 1 in 5 deals where the buyer explicitly confirmed AI discovery. After the metadata layer was in place (Month 3), we could trace those AI touchpoints back to specific content pieces and keyword clusters, confirming which content was generating which pipeline. After the refresh flywheel stabilized (Month 6), we had trend data: AI-influenced deals were closing moderately faster than non-AI deals and at comparable or slightly larger deal sizes — consistent with the industry pattern documented by Yolando (ChatGPT-sourced leads: 20 percent more revenue, 40 percent faster close).

Content efficiency: Our 50-article blog, produced manually, cost approximately $25,000 per year in creation and refresh labor. Under the agent-automated model with scheduled refresh, the equivalent program runs at a fraction of that cost — the metadata layer that makes attribution possible is a byproduct of the agent architecture, not an additional expense.

The gap ratio: Our self-reported AI attribution (20 percent of qualified opportunities) versus our analytics-tracked AI referral traffic (approximately 0.4 percent of sessions) suggested a gap ratio of roughly 50× — consistent with the industry range documented in the 54x gap diagnostic. We weren't unique. We had the same measurement problem every GEO program has. The difference was that we had built the instrumentation to measure it.

What we couldn't yet measure: At the six-month mark, our dataset was directionally informative but not statistically robust. With roughly 50 to 60 qualified opportunities per quarter and AI touchpoints confirmed on 20 percent, we had about 10 to 12 AI-influenced deals per quarter. Trend lines were visible; statistical significance was not. The architecture was proven. The sample size needs more time.

What We'd Do Differently

Three things, with the benefit of hindsight:

1. Instrument CRM before creating a single piece of GEO content. We spent six months producing GEO content with zero attribution infrastructure. Every article published in that period is a missed measurement opportunity — we'll never know whether those articles generated pipeline because the CRM wasn't instrumented when they were earning citations. The correct sequence is: CRM fields → form fields → SDR script → THEN content production. Not the reverse. We knew this intellectually from the 4-layer architecture. We just didn't follow our own framework.

2. Start with commercial-intent queries, not informational ones. Our early GEO content targeted high-volume informational keywords like "what is generative engine optimization?" and "GEO vs SEO." These articles earned citations — our dashboard looked healthy. But when we instrumented the CRM, we discovered that the citations actually generating pipeline came from much lower-volume, higher-intent queries: "best GEO software for agent automation," "GEO tool with CRM integration," "attribution-native content platform." We had been optimizing for the wrong thing — citation volume instead of pipeline connection. The Intent-Segmented GEO strategy we now recommend to customers is based on this mistake.

3. Treat the SDR script as product infrastructure, not sales enablement. We initially positioned the AI discovery question as a "nice to have" in our sales script. Compliance was inconsistent — some SDRs asked it every call, some forgot entirely. It took two months to realize that the SDR script is the most important data collection mechanism in the GEO attribution stack. Every missed question is a deal whose AI influence goes unrecorded. We rebuilt the script as a required field in our CRM — the opportunity can't advance without an AI Discovery Source value — and compliance went to 100 percent overnight.

What You Can Take Away

The $0 fix is real. Our CRM instrumentation cost nothing and surfaced more attribution signal than any paid tool. If you do one thing after reading this case study, add the AI Search lead source field and the "How did you hear about us?" free-text field to your CRM today. The 54x Gap Diagnostic walks through implementation step by step.

Attribution-native content is an architecture decision, not a content quality decision. The difference between our pre-instrumentation and post-instrumentation GEO programs wasn't that the content got better. It was that the content carried metadata that made attribution possible. Agent automation made this economically viable — manual metadata tagging at scale is unsustainable, as detailed in the Attribution-Native Content article.

The sequence matters more than the tools. The 4-Layer GEO Measurement Architecture we built — CRM Ground Truth → Branded Search Lift → Entity Association Monitoring → Citation Distribution Sampling — is an investment sequence, not just a reporting framework. Implement Layer 1 ($0) before Layer 4 ($295/month). We made the mistake of monitoring citations before instrumenting CRM. We don't recommend repeating it.

The sample size problem is real. Six months of data with 10 to 12 AI-influenced deals per quarter produces directionally useful trends, not statistically significant conclusions. This is not a failure of the architecture — it's a function of B2B deal velocity. The companies that build this infrastructure now will have statistically robust GEO attribution data in 12 to 18 months, precisely when the CFO starts demanding it. The window to build before the question is asked is now.

Siteup's agent automation platform embeds funnel-stage, keyword-cluster, and conversion-endpoint metadata into every piece of content at creation — the architecture described in this case study. If you want to build an attribution-native GEO program for your own brand, sign up for Siteup or see our pricing.