RAG Pilot Failures: Why Your AEO Budget Splits in Two

The Revenue Signal · Issue 07

In February, Kevin Indig, the growth advisor who led SEO at Shopify, G2, and Atlassian, published research that should have changed how every B2B company invests in AI visibility.

Indig and his team analyzed 3 million ChatGPT responses, 30 million citations, and isolated 18,012 verified citations. They used sentence-transformer embeddings to match each AI response back to the source sentence the model extracted from. Then they measured where in the source content that sentence appeared.

The finding: 44.2% of all ChatGPT citations come from the first 30% of a page. The result held across randomized validation batches. Indig called it "statistically indisputable."

What that means in operational terms: the engineering decision that determines whether ChatGPT cites your content is the same engineering decision that determines whether your internal AI agent retrieves the right answer when a customer asks it a question. Similar retrieval problem. Similar content requirements. Overlapping failure modes.

Most B2B companies are paying two separate teams to solve this. Marketing owns AEO. Engineering owns RAG. The CFO sees two AI line items and does not know that the work underneath is one problem.

This issue covers the unified retrieval problem — why 80% of enterprise RAG pilots never reach production, why AEO citation rates are flat for most B2B SaaS, and why both failures share a root cause. The Signal is the data. The Build is Mentimeter. The Move is what to do this week. Reading time: about 12 minutes.

AI answer engines and internal RAG agents share a similar four-stage retrieval pattern. Content that fails one tends to fail the other.

Most AI answer engines and internal RAG agents follow a similar pattern. They retrieve candidate passages based on the query, rank them for relevance and trust signals, synthesize an answer from the top-ranked passages, and attach citations when citations are supported.

This is not just a marketing analogy. Most AI answer engines behave like RAG pipelines pointed at the public web, with different architectures but similar mechanics. Your internal customer-support agent is a RAG pipeline pointed at Confluence and your help docs.

The failure data tells the rest of the story.

Industry analyses commonly place enterprise RAG failure rates in the 70%–80% range, though the exact figure varies by definition and source. RAG About It's March 2026 analysis cites this band. SSON Intelligent Automation Week conference data states 80% of enterprise RAG projects experience critical failures, with only 20% achieving sustained success. The global RAG market is on track to exceed $40 billion by 2026, growing fastest where the failure rate is highest.

VentureBeat's Q1 2026 enterprise survey shows enterprise intent to adopt hybrid retrieval tripled from 10.3% in January to 33.3% in March 2026 — even as 22% of qualified respondents reported no production RAG systems at all. VentureBeat's read: the market is shifting from adding retrieval layers to fixing the ones already in place.

The AEO side looks identical. The found that 89% of domains cited by at least one major AI engine are cited by only one. Most B2B companies are visible in one LLM and invisible in the other three. BrightEdge data covering February 2025 to February 2026: the share of B2B tech queries that trigger an AI Overview grew from 36% to 82%. The default search experience for technical buyers is now AI synthesis, not blue links.

The CEO read is uncomfortable. If the same engineering decisions determine outcomes on both surfaces (chunking, schema, freshness, authority signals, retrieval grading), then the company running two budgets is running two lower-performing programs against the same root cause. The CFO who consolidates wins twice. The CFO who does not consolidate pays twice and underperforms in both.

Indig's finding makes the consolidation case operational. Front-loading key answers in the first 30% of a page is supported by his data for external AI citation. The same structure tends to help internal RAG systems that chunk and rank documents from the top down. The same content rewrite ships in both directions.

How Mentimeter restructured comparison content for AI extraction — and why the same pages serve both customer support and ChatGPT.

Mentimeter, the interactive presentation platform used by hundreds of millions of people across business and education, faced the version of this problem most B2B SaaS companies are about to recognize as their own. Buyers were asking AI engines to recommend presentation tools. Mentimeter was inconsistently cited. The fix surfaced an insight worth more than the fix itself.

Working with Siege Media, the team restructured comparison content and feature documentation around three principles:

Dual audience targeting. The same comparison page served two distinct buyer personas — educators evaluating classroom tools and business teams evaluating presentation software. Rather than build separate content tracks, the team produced single pages with structured sections that AI engines could extract for either query type.

Comparison-focused structure. Pages directly answered "versus" queries that AI systems commonly receive. The structural choice was deliberate — comparison tables extract cleanly from RAG pipelines because they encode multiple data points in a format that maps to how buyers prompt LLMs.

Feature-specific organization. Content was reorganized around user needs rather than product taxonomy. A feature page on "live polling for hybrid meetings" beat a feature page titled "Polls" for retrieval purposes because it matched the language buyers actually use when prompting an AI engine.

The transferable insight is in what did not get built: a separate set of content for AI engines. The same comparison pages and feature documentation that earned external ChatGPT, Perplexity, and Claude citations also serve as the trusted source the customer support team's internal AI assistants retrieve from when a paying customer asks how to set up a specific feature. One asset class. Two retrieval surfaces. One ownership team.

Mentimeter earned AI citations on competitive comparison queries — capturing buyers at the decision stage of their research. The strategic lesson is structural, not tactical. The companies winning AEO in 2026 are the ones who recognized their help documentation and comparison content already had to serve internal AI agents, and structured it to serve both at once.

Depth of structure beats volume of content. The same retrieval discipline applied to both surfaces wins both.

Three actions before next Thursday.

Step 1 takes a Tuesday afternoon. The full diagnostic version is the Revenue Experts AI Citation Audit Method: 50 buyer-intent prompts × 4 LLMs × 3 runs = 600 measured calls, human-delivered in 5–7 business days, $497. It returns a numerical citation score, a competitor map, and a gap-to-action matrix — the artifacts your CFO needs to defend the team consolidation in step 3.

Book the Citation Audit →

Elizabeta Kuzevska Co-Founder, Revenue Experts AI https://revenueexperts.ai