Study
The State of AI Search Visibility 2026
AI search visibility benchmarks get published faster than the measurement standards behind them. I still want the order reversed.
The State of AI Search Visibility 2026 is a DankRank AI SEO research study designed to compare brand and source visibility across ChatGPT Search, Gemini, Perplexity, Google AI Mode and traditional Google Search.
The methodology is locked. The final benchmark collection is not underway.
01
Research Status
Methodology v1.2 locked. Final benchmark collection deferred.
There are no aggregate findings to publish yet.
I completed a 20-observation operational pilot to test whether the collection method actually worked in the live products. Pilot observations are permanently excluded from the benchmark. The pilot exposed a few coding problems around local result modules, first-party citations and source classification, which is exactly the kind of thing I wanted to catch before collecting the real dataset. The rules were corrected and independently checked before I locked v1.2.
If final collection resumes, the study will contain 176 observations built from 32 exact prompts. I am not treating the pilot as benchmark evidence, and I am not filling this page with percentages borrowed from another vendor's study while my own final dataset does not exist.
02
The Questions We're Testing
The study is built around a few questions I think are more useful than a generic "who wins AI search?" ranking.
How often do the same sources appear across different AI systems?How often is a brand mentioned without its own site being cited?How much do recommendations change when the prompt moves from informational to buyer intent?Which source types appear repeatedly?How stable are the observed results when selected commercial prompts are repeated?How different is the source set from traditional Google results for the same underlying topic?
Total AI market share, consumer adoption and conversion sit outside this study's scope.
03
Why 32 Prompts and 176 Observations
I originally published a 240-prompt design. Before any final-study data was collected, I cut it back.
The old plan would have produced far more observations than I could justify for the questions this study is meant to answer. The locked v1.2 design keeps the breadth I care about without turning an agency research project into a university-scale benchmark program.
The prompt set is US-primary and covers eight topic strata. Each stratum contributes two paired topics, and each pair has one informational prompt and one commercial prompt. Together, the pairs give the study 16 informational and 16 commercial prompts.
Each prompt runs once across four AI products, creating 128 primary AI observations. Ordinary Google is collected for 16 comparator prompts. Eight commercial prompts are then repeated across all four AI products 48 to 72 hours later, adding 32 repeat observations.
Total final-study scale: 176 observations.
The earlier 240-prompt design remains part of the research history, but it is not the method that will produce the benchmark if collection resumes.
04
What We Record for Each Observation
For every prompt and environment, the dataset is designed to preserve the raw answer or result set before it is coded.
The record includes:
engine or productprompt and prompt IDobservation date and timevisible model or mode when the product exposes onebrand mentionsexplicit recommendations or consideration setscited domains and source URLssource typefirst-party citation presencelocal-module website destinations where a product shows business cards without editorial citationsrank or order where the interface meaningfully exposes onetraditional Google organic results for the paired comparator samplechanges across the repeated commercial observations
I preserve the underlying answer and source evidence before turning it into any aggregate rate or comparison. There is no single blended "AI visibility score" in this study.
05
Citation Visibility + Recommendation Visibility Stay Separate
A source can be cited without the associated brand being recommended. A brand can also be recommended while a different domain supplies the supporting citation.
The benchmark keeps those behaviors separate. It also distinguishes first-party citations from third-party support instead of treating every brand mention as the same kind of visibility.
06
Source Overlap Is One of the Most Interesting Cuts
I want to know how often different AI systems rely on the same sources. Recurring publishers, forums, first-party sites and reference sources tell us something useful about the information around a query. Low overlap points the other way: the products are drawing from different source sets.
The study will report source overlap only from the observed prompt set. It will not turn a 32-prompt sample into a universal statement about every query on an engine.
07
Buyer Intent Gets Its Own View
An informational question and a provider-selection question can produce very different answer patterns. The paired design lets me compare those two jobs without blending them into one rate.
The distinction matters because a product can be useful as an information source without ever becoming a recommendation when the query moves closer to a buying decision.
08
What the Traditional Google Comparison Can and Can't Tell Us
For 16 of the 32 prompts, the study also records the ordinary organic results available on the first Google results page, up to rank 10.
If Google exposes fewer than 10 ordinary organic results, I record what is actually there. I do not pad the set with Maps, shopping, AI modules or results from a second page just to make the number reach 10.
The comparison can show whether the same domains, brands or page types appear across Google and the AI products. It cannot establish that one system copied the other or that an organic ranking caused an AI citation. Correlation stays correlation.
09
Repeated Observations Matter
AI responses can change between runs. A single prompt snapshot tells me what happened at that moment. It is much weaker evidence for a durable recommendation pattern.
The final design repeats eight commercial prompts across all four AI products 48 to 72 hours after the primary run starts. The 32 repeat observations are analyzed separately from the primary rates so volatility stays visible instead of being averaged away.
10
Limitations We Already Know About
The prompt set is a sample, not the entire search universe.AI products change quickly.The products do not expose citations or sources in exactly the same way.Interfaces and answers can differ by account, geography, model or product state.Repeated outputs may vary.A citation does not establish business influence or conversion. A brand mention, by itself, says nothing about source authority.The study observes behavior from the outside and cannot see proprietary retrieval or ranking systems.
Citation availability will be reported by environment as well as pooled across the four AI products, because a zero-citation answer and a product that simply exposes sources differently are not the same measurement problem.
If final collection resumes, the collection record will disclose the actual collector geography and the visible account/product state used for each environment. If an account carries prior history materially related to the study, that gets disclosed too. I am not going to manufacture a synthetic location and call it a universal answer.
The limitations change how the findings can be phrased. They do not disappear because a percentage looks clean.
11
What Is Published Today
Right now, this page is the public research record for the locked methodology and its limitations. It is not a results report.
Published now:
methodology v1.2 statusstudy questionsfinal-study sample designmeasurement distinctionsknown limitationsrelated measurement and evidence resources
Not published because it does not exist yet:
final benchmark datasetaggregate visibility findingssource-overlap findingsbuyer-intent resultsengine-by-engine benchmark ratesrepeatability resultsany claim about market share, consumer adoption or conversion
If final collection is authorized later and the dataset passes independent QA, this page can be rebuilt around the actual findings. Until then, the honest version of the article is the methodology record, not a pretend benchmark.
12
Related Measurement + Evidence
For DankRank's own new-domain starting point, see the Launch AI Visibility Baseline.
For the practical mechanics behind prompt sets, citations, competitors and reporting, use How to Measure AI Search Visibility.
For why the baseline comes before the intervention, read Measure AI Visibility Before You Optimize It.
Back to the Research hub for the full research program.