Recorded
Directly logged inside a named instrument, within that instrument's stated coverage.
Public methodology, version 1.4
How Sequoia GEO measures AI search without turning visibility into a lead
AI answers change. Citations are not recommendations. Referral visits are not inquiries. This page defines every stage before Sequoia reports it.
Sequoia built the AI Search Evidence System used inside its AI SEO engagements. The current release combines dated manual question panels, public-source reconciliation, lead attribution, and the evidence stages defined below. An automated OpenAI query and source collector is still pilot-only until its first live observation is validated. The system is included in the engagement, not sold as a separate software subscription.
Published August 20, 2026. Last revised August 30, 2026.
This AI-specific method operates under the broader Sequoia Proof System, which governs evidence classes, corrections, verification, causation, and publication limits across all client work.
Confidence and coverage
How confident can an AI visibility assessment be?
We document where your business appears, how consistently it appears, which sources support the answer, and what the evidence can actually prove. A controlled prompt panel measures the conditions we record. It does not represent every question, user, location, or AI session.
Observed
Preserved under a documented prompt and session protocol on a stated date.
Inferred
Supported by evidence only when the assumption connecting the evidence is stated.
Unknown
No available instrument or identity-level join can verify the claim.
Generated-answer evidence
Eligibility
Citation
Mention
Recommendation
Measurement gap
Most AI products do not preserve an identity-level join from a generated answer to the person who later visits or contacts the business.
Business-outcome evidence
Identifiable referral
Inquiry
Qualified lead
Job booked
Sequoia reports both chains, but does not fill the gap with an assumed attribution. A change is described as movement only when the comparison uses a frozen panel and an appropriate reference, such as untreated pages, reference prompts, or a no-intervention arm.
The reporting ladder
Eight stages, eight different claims
A result can move through several stages, but it is counted once at each stage and never promoted without the required evidence.
This page applies Sequoia’s own pipeline qualification rule as a business-specific instance of the general qualification rule in the Sequoia Proof System.
StageWhat it meansWhat it does not prove
Mentioned
The business appears anywhere in the response.
That the assistant recommends hiring the business.
Recommended
The response presents the business as a plausible provider to hire.
That the business is the first or strongest recommendation.
Primary recommendation
The response presents the business first or as the clearest recommended provider.
That the position will repeat for another user or another run.
Cited
A page from the business is shown as a source in the answer.
That the business itself was mentioned or recommended.
Referral visit
A captured website session arrives from an identified AI source or registered campaign link.
That the visitor became an inquiry or that every AI referral was captured.
Inquiry
A real person contacts the business about a possible service need.
That the business, need, authority, or source meets the qualification rule.
Qualified lead
A real business, an addressable Sequoia GEO need, and a decision-maker response or meeting are established.
That the opportunity will become a customer or job booked.
Job booked
The operating business confirms a booked job under its normal written definition.
That a prior citation or prompt observation caused the booking.
Controlled prompt observations
How an observation wave is run
The goal is a repeatable sample with known limitations, not a screenshot selected because it looks favorable.
01
Freeze the question set
Pre-register a frozen core of commercial and control questions, plus a separately reported rotating set. Record the geography and inclusion rules before the wave begins. Do not rewrite questions after seeing the answers.
02
Control the session
Use a fresh session with memory disabled where the product permits it. Record the platform, visible product, app or web surface, account state, search or browse state, device, geography, location method, and date.
03
Repeat each observation
Run each frozen question five times in separate fresh sessions per platform. The prompt-platform cell is the analysis unit. The repeats measure answer stability and are not treated as statistically independent trials.
04
Archive before coding
Preserve the complete response privately with the timestamp, visible citations, companies named, and search state. Code the observation only after the artifact is saved.
05
Code each stage separately
Record mentioned, recommended, primary, cited, or absent independently. A cited Sequoia page does not automatically create a Sequoia recommendation.
06
Check coding reliability
Have an independent second reviewer code at least 20 percent of observations before drawing a conclusion. Preserve both decisions, report disagreements by field, and stop publication when a material code cannot be resolved against the public definition.
Variance and limits
AI answers are observations, not universal rankings
Results can vary by platform, product version, prompt wording, geography, account state, memory, retrieval behavior, and date. Repeated observations reduce the risk of overreacting to one answer, but they do not create a market-wide ranking.
Multiple runs of the same question on the same platform are correlated observations. Sequoia uses those repeats to describe stability, not to inflate the sample size or imply statistical independence.
Sequoia reports platforms separately and preserves misses as well as appearances. Results are expressed as observed counts or shares within the frozen sample, never as universal search volume.
A 30-day review can identify implementation failures. Material recommendation changes are evaluated over 60 to 90 days after discovery and indexing, with no guarantee that a correction will change an independently controlled answer.
Required context for a published number
- Business or property measured
- Platform, account report, and metric definition
- Exact date range and read date
- Prompt panel, frozen or rotating arm, repetitions, and session controls
- App or web surface, account state, device, geography, and location method
- Whether the result is observed, self-reported, reconstructed, or unavailable
- Explicit statement of what the metric does not establish
Referral evidence
Analytics can capture some assistant referrals and registered tracking links. Direct calls, copied links, privacy controls, and untagged journeys can leave the source unknown.
Reported recommendations
A prospect's account of how they found Sequoia is valuable intake evidence. Without the original artifact, it is not proof of the exact prompt, answer, or causal source.
Commercial outcomes
Inquiries, qualified leads, meetings, and jobs booked are reconciled in the operating system. A citation or scheduled calendar event cannot be substituted for qualification.
Publication rules
What Sequoia commits to publishing honestly
- ✓Publish misses and absences alongside favorable observations.
- ✓Do not imply that a correction caused a recommendation without a defensible design and observation window.
- ✓Keep customer identity, artifacts, and outcomes private unless written permission covers the specific publication.
- ✓Separate owned-page citations from independent corroboration.
- ✓Preserve the frozen question set for the reporting period and date later additions separately.
- ✓Update definitions through a visible changelog instead of silently rewriting prior results.
Changelog
August 30, 2026
Version 1.4 added explicit confidence labels, documented the gap between generated-answer evidence and identifiable business outcomes, and expanded prompt-panel and session controls.
August 29, 2026
Version 1.3 placed this channel-specific method under the Sequoia Proof System and clarified how Sequoia’s pipeline qualification rule relates to the general proof standard.
August 21, 2026
Version 1.2 documented Sequoia’s included AI Search Evidence System, identified the automated collector as pilot-only pending live validation, and clarified that the system is delivery infrastructure rather than a separate software product.
August 21, 2026
Version 1.1 clarified that repeat runs within a prompt-platform cell measure stability and are not statistically independent trials.
August 20, 2026
Version 1.0 published with stage definitions, five-run observation protocol, variance rules, source context, privacy boundaries, and publication commitments.
See what this method looks like when applied
Read the current Sequoia visibility snapshot or request a public-surface assessment for your business.