Still Fox Free snapshot

Measure, fix, measure again.

Still Fox measures how AI assistants name and describe businesses, using repeated runs on the assistants' official search APIs, from Seattle.

An animated field of cells flows from left to right like a river. At a vertical reading line the cells snap into ordered rows that look like the lines of a page, and the page scrolls slowly upward as reading continues. One cell in the page, outlined, stands for the business.

Evidence ledger

This page is the evidence behind what Still Fox sells and what it refuses to sell: every tactic is graded by the strength of the study behind it, and our own measurements are designed in writing before any data comes in. If anyone has pitched you AI search work, hold them to the same standard: ask for their grades, their intervals and the list of what they refuse to sell.

Each tactic below is graded by the strength of the study behind it, including the ones we refuse to sell. Last updated 2026-10-11.

How to read a grade. A grade rates how much to trust the study, not how much the tactic will help you. A is a large sample or a controlled test; D is a claim with no data shown. A row tagged vendor comes from a company that sells these services.

Grades

A
A large sample (thousands of units) or a controlled test
B
Several consistent studies, or one mid-size study with a disclosed method
C
One small study, a single case, an expert survey or one firm's reverse-engineering
D
An assertion without disclosed data, or a number seen only secondhand
EC
The engine operator's own documentation: describes mechanism, never effect size

Qualifiers

corr
A correlation, not a controlled test
vendor
Published by a company that sells GEO, SEO, listings or marketing services
null
The study looked for an effect and found none
Tactics, their evidence and our position, as of 2026-10-11
Tactic Verdict Evidence Grade Sources Our position
Open AI crawlers and server-rendered pages Required Blocking OAI-SearchBot removes a site from ChatGPT search answers. Major AI crawlers did not execute JavaScript in 2024 tests. EC; C, dated Built into every rebuild and hardening
One page per service and location Plausible Practitioners rank it the second AI factor. Businesses that assistants found average 34.8 pages against 15.2 for missed ones. C Built into every rebuild
Accurate listings on Business Profile, Yelp and Bing Places Needed; effect on selection unmeasured ChatGPT and Perplexity were about 68% accurate on contact data. 55.58% of ChatGPT results fully matched the Business Profile. B, vendor Checked in every snapshot and care plan
Rank in Google and Bing for the query and its sub-questions Works Page-one Google rank correlates about 0.65 with ChatGPT brand mentions across 10,000 answers. 71.1% of the most-recommended businesses rank for their own keywords against 13.5% of the least. A, corr, vendor The agency's SEO work; our builds remove the technical blocks
Appear on third-party "best of" lists Works "Best X" lists were 43.8% of cited page types across 750 prompts. 47 practitioners rank curated lists the top AI factor. B The snapshot's source map shows which lists the assistants cite
Earn brand mentions on other sites Works Across 75,000 brands, branded web mentions correlate 0.664 with ChatGPT visibility (Ahrefs, December 2025, three surfaces). In an earlier Ahrefs study (May 2025, AI Overviews), branded mentions correlated 0.664 while the number of backlinks correlated 0.218 and referring domains 0.295. A, corr, vendor The source map shows where mentions count
Review volume and rating Works The most-recommended tier averages 304.2 reviews against 117.5. Recommended locations average 4.3 stars. A, corr, vendor The agency's reputation work
Answer-first pages and Q&A blocks Unproven Microsoft recommends them. Google says no chunking is needed. EC, conflicting Not sold as a lever
YouTube presence Promising YouTube mentions correlate about 0.737 with AI visibility, the strongest factor across 75,000 brands. A, corr, vendor Not offered
Schema markup as a citation lever No measured AI effect 1,885 pages against 4,000 controls: plus 2.4% in AI Mode and plus 2.2% in ChatGPT (not significant), minus 4.6% in AI Overviews. A, causal null Included in every build, never sold
Rewriting pages with quotations and statistics Mostly ineffective on live engines A lab study reported gains up to about 40%. A NeurIPS benchmark found most such rewrites ineffective or harmful and traditional SEO more effective; a survey of 45 studies reports 3 of 54 method and domain combinations significantly positive. A in the lab; A independent; B secondhand Refused
llms.txt files Folklore No correlation with citations across about 300,000 domains. 97% of files got zero requests in May 2026. Google Search ignores it. A, null; EC Refused
Reddit seeding Unreliable Reddit pages retrieved by ChatGPT were cited 1.93% of the time across 1.4 million prompts. A, vendor, 2025 data Refused
Self-ranked "best of" lists Probably works; inauthentic A self-published list sat behind 34% of ChatGPT software responses that recommended a brand from its own site. Google calls inauthentic mention-seeking unhelpful. B, vendor; EC Refused
"Summarize with AI" memory-injection buttons Avoid Microsoft found 50 or more such prompts from 31 companies in 60 days and deployed mitigations. EC Refused

Pre-registered baseline

Status: the baseline is not yet published. The prompts and this design were frozen on 2026-10-11. Any change after the first real run starts version 2, and the change is recorded.

Question
Does an AI-ready own site (build number one) move the share of AI-assistant answers that name Still Fox on niche queries, compared with a broad control cluster?
Design
  • Prompts: a fixed set in three groups. Two groups are niche questions, one a partner agency would ask and one from our first vertical (see the index), and the third is broad control questions.
  • Engines: the official search APIs of OpenAI, Perplexity and Anthropic. Google Gemini is pending its API terms on storing grounded answers, and joins only if they allow it.
  • Runs: each prompt is asked many times per engine per window. The location Seattle is in the prompt text where it matters and is also sent as a location parameter where the API supports it.
  • Surface: API only. Consumer-app checks are manual, on a fixed small sample, and reported separately.
  • Windows: a baseline before this site launches publicly, then every 30 days.
Primary measure
The share of answers that name the target (Still Fox or an alias), per cluster and per engine, with 95% Wilson intervals.
Expectations, stated before data
  • The baseline names Still Fox in none of the answers, on every cluster. Even then, the 95% upper bound on the true share stays above zero, so a result of no mentions is a bound, not a proof.
  • Detecting a move from 10% to 20% takes about 200 runs per arm, so a single-window change inside one engine is not a claim.
Decision rule
On day 180 of the project, if every niche cluster is still at 0% with inputs complete, we drop the "recommended" claim.

Every result from this design publishes on the Research page, whatever it shows. The first entries are in preparation.

Measure, fix, measure again

  1. Measure. A fixed set of prompts, each run several times on each assistant's official search API. Every answer and every cited page is stored with its date, model and cost.
  2. Fix. Crawler access at robots and CDN, server-rendered pages, one page per service and office, the same facts everywhere.
  3. Measure again. The share of answers that name the business, with a 95% interval, before and after, plus every factual error that remains.

Why repeated runs

Assistants give different answers to the same question. One answer is an anecdote. Twenty runs with an interval is a measurement.

What we skip

Tactics without evidence behind them. No llms.txt files, no structured data sold as a ranking lever, no statistics stuffing, no self-ranked "best of" lists, no Reddit seeding. The ledger above gives the evidence for each.

Claims and sources

“With twenty runs, an assistant that never names the business still has an upper bound near 16%”

Wilson score interval, 0 of 20, 95%. With z = 1.96, the upper bound for 0 of n is z² / (n + z²), which is 16.1% for n = 20.

“Detecting a move from 10% to 20% takes about 200 runs per arm”

A standard two-proportion sample-size calculation, two-sided test at 5% significance with 80% power, which gives 199 per arm.

“The evidence graphic on the home page shows 19 grades on 15 tactics, 8 of them on tactics we refuse to sell”

Counted from the evidence ledger above, one mark per grade a tactic carries (a row graded EC; C, dated carries two). Grade A has 8 marks, B 4, C 2, D 0 and EC 5, which makes 19. The five tactics marked Refused carry 3 A marks, 2 B marks and 3 EC marks. Each mark is drawn as 7 dots, and dot size follows the largest sample the row’s evidence text cites.

“The interval graphic on the home page is an illustration, not a result”

It replays 20 made-up runs, 11 of which name the business, and draws the 95% Wilson interval of the running tally. At 11 of 20 the interval is 34.2% to 74.2% (z = 1.96).

  • API answers can differ from the consumer apps. Consumer apps are checked by hand on a fixed sample and reported separately.
  • AI answers can contain factual errors. Snapshots list them and do not vouch for them.
  • The ledger is transcribed from our research notes, with the wording shortened and the numbers and grades unchanged. Each row links to its sources; check any figure there. A grade describes the strength of the study, not the size of the effect, and a row marked vendor comes from a company that sells GEO, SEO, listings or marketing services.
  • The expectations above were stated before data.

Get a reading of what the assistants say about your business.

The snapshot is free. It shows whether AI crawlers can reach your site and what the assistants say about you, by email in two business days.

Request a free snapshot For agencies