Why this trial
No one has yet shown, with random assignment and a control group, that optimizing for AI search changes what ChatGPT, Perplexity, Google AI Mode and Copilot recommend. This trial tests it with 10 to 15 software and ecommerce brands, and the plan is published before any data are collected.
Observational studies show what AI assistants cite: Ahrefs on Google overlap, Omnipresence on local recommendations, and Scrunch on self-ranking lists. The July 2026 survey of 45 studies found no stable, long-term, cross-platform causal effect on discoverability or downstream behavior. Our own 100-answer ChatGPT study is observational too. This trial uses random assignment instead.
The question
Does publishing an answer-first page with original, specific evidence for a buyer topic make ChatGPT, Perplexity, Google AI Mode and Microsoft Copilot mention, recommend and cite the brand more often for that topic? Does it send more visitors from those assistants?
Design
- Brands. 10 to 15 B2B software or ecommerce brands, about 20 to 500 employees, in English-language markets.
- Topics. 20 buyer topics per brand, agreed before randomization and chosen so they do not overlap.
- Randomization. Topics are paired by baseline mention rate. A computer coin flip within each pair assigns one to treatment and one to control, using a seed published with the registration.
- Treatment. One new or rewritten page per treated topic. It answers the question first, then gives evidence the brand can stand behind, such as real prices, test results or customer data. It follows Google's spam policies: no self-ranking lists and no scaled pages.
- Control. No new page, rewrite or promotion for control topics for 12 weeks. Afterwards they get the same treatment.
Outcomes
| Outcome | Definition | Role |
|---|---|---|
| Brand mention rate | Share of prompt runs per topic where the brand is named in the answer | Primary |
| Recommendation rate | Share of runs where the brand is in a suggested shortlist | Secondary |
| Citation rate | Share of runs citing any page on the brand's domain | Secondary |
| Target-page citation | Share of runs citing the treated page itself | Secondary |
| AI referral sessions | Analytics sessions from ChatGPT, Perplexity, Copilot and Gemini to the topic's pages | Secondary |
| Description accuracy | Share of mentions that describe the brand correctly | Guardrail |
Each topic has 3 fixed prompts that do not name the brand. Each prompt is asked twice per engine, twice a week, in the assistants' consumer apps rather than their APIs, because the API cites different pages from the app. A person checks 10% of the automatic scores.
Analysis
- Primary estimate. The treated-minus-control difference in mention rate within each topic pair over follow-up weeks 5 to 12, adjusted for the pair's baseline. Pair differences are averaged within each brand, then across brands, and the confidence interval comes from the spread between brands.
- Per engine. Each engine is reported separately, with Holm-adjusted p-values across the four.
- Power. Our simulation suggests about 80% power to detect a 5-point rise in mention rate with 10 brands. The baseline will check the simulation's assumptions.
- No early stopping. The readout is at week 12.
- Publication. The result is published whatever it shows, with the raw answers, citations and scores. Brands can stay anonymous.
Timeline
| Step | When |
|---|---|
| Recruit brands | Oct 26 to Nov 13, 2026 |
| Register this plan on OSF | Nov 16, 2026, before any baseline data |
| Baseline measurement | Nov 16 to Nov 29, 2026 |
| Randomize topics | Nov 30, 2026 |
| Write and publish treated pages | Nov 30 to Dec 18, 2026 |
| Follow-up measurement | Dec 18, 2026 to Mar 12, 2027 (12 weeks) |
| Readout and publication | Late March 2027 |
| Control topics treated | Mar 15 to Apr 9, 2027 |
Limits and disclosure
Results will apply to B2B software and ecommerce brands in English-language markets during the trial period. They will not show what works for local businesses or other languages. NRLC.ai designs the trial, writes the briefs and runs the analysis. That is why the plan is fixed in advance and the data will be public for anyone to re-analyze.
Join the trial
We are looking for 10 to 15 brands that already track AI visibility and have a content team. Book a call or email joel.m@neuralcommandllc.com with "AI search trial" in the subject.
FAQ
Has anyone tested AI search optimization with a randomized trial?
Not that we can find. A July 2026 survey of 45 generative engine optimization studies concluded that no reviewed technique shows a stable, longitudinal, cross-platform causal effect on discoverability or downstream behavior. Published studies are observational.
What does a participating brand get?
20 buyer topics measured in ChatGPT, Perplexity, Google AI Mode and Copilot for 14 weeks, and 20 pages briefed and checked by NRLC.ai: half during the trial and half after week 12. Each brand sees its own results before publication and can stay anonymous.
What if the result shows no effect?
It is published anyway. The plan, primary outcome and analysis are fixed before data collection, and any later change is listed as a deviation.
Why randomize topics inside each brand instead of comparing brands?
Brands differ in size, authority and reputation. Splitting each brand's own topics into treated and control groups makes each brand its own comparison, so those differences cancel out.
Version history
- October 23, 2026: first public version.