A practice that builds and runs the site the assistants read.
Still Fox is a Seattle practice that builds and runs websites AI assistants read and measures what they say about a business.
What the practice builds and runs
The practice builds and runs the website that AI assistants read about a business: reachable, accurate, measured every month, and owned by the business from day one. Reachable means every search and user crawler is let in at the robots file and at the firewall, and the text is in the HTML. Accurate means one record of the business's facts sits behind every page, and the errors the assistants make are counted against it each month. Measured means the share of answers, with the run count, the surface and the interval beside every number. Owned means the domain, the content and the repository are in the business's name from the first day.
We refuse levers that cannot be measured, guarantees of rankings, mentions or leads, measurement sold alone, dashboards sold in place of the work, and lock-in.
We rebuild a site, harden the one a business already has, and run it on a monthly care plan. Every page opens with who the business is, what it does and where. The pages are server-rendered HTML with no client JavaScript by default, the crawlers are allowed in robots.txt and at the CDN (the firewall layer in front of the site), and structured data is included in every build and never sold. Agencies: we build and run the same work wholesale, under your name, and you keep the client.
What the assistants read
When someone asks an AI assistant such as ChatGPT, Perplexity, Gemini or Claude who to call, the assistant answers in sentences, not a list of links. It describes a business from what it can reach and read, and the business's own website is the one source the business controls. A site fails the assistants in plain ways: robots.txt or the firewall turns their crawlers away, the text appears only after a script runs, or the name, address and phone differ from page to page. Most of these never show up when the owner opens the site in a browser.
What we measure
A fixed set of prompts runs several times on each assistant's official search API, and we store every answer and every page it cites. We report the share of answers that name the business, with a 95% interval, before and after the fix, and on a care plan again every month. We list every fact the assistants get wrong. Our own panel asks a fixed set of buyer questions, many times each, on the official APIs of OpenAI, Perplexity and Anthropic. Gemini joins when its API terms allow us to store the answers. The Method page shows the design.
How the work gets done
Our tooling runs the checks and the measurement, and drafts pages from the client's own material. The practice reads every page before it goes to the client. The client approves every page before it goes live. For a law firm, a lawyer signs off after that.
The split is deliberate. Software repeats a check the same way every time, so we let it run the checks. A sentence on a client's site needs a reader who answers for it, so a person reads it.
What we refuse to sell
Some tactics have no evidence behind them, or a negative result. We do not sell them, and the reason is on the record.
- llms.txt files
- We do not sell them. A large study found no link between the file and citations, and Google Search ignores it.
- Structured data as a ranking lever
- We do not sell it as one. A controlled test found no measured effect on AI citations. We include structured data in every build at no charge.
- Statistics stuffing
- We do not rewrite pages to pile in quotations and statistics. One lab study reported gains. An independent benchmark on live engines found most such rewrites ineffective or harmful.
- Self-ranked "best of" lists
- We do not publish lists that rank a client first. The tactic may work, but it is inauthentic, and Google calls inauthentic mention-seeking unhelpful. A practice that grades evidence cannot rank itself.
- Reddit seeding
- We do not seed threads. In a vendor study, the Reddit pages ChatGPT retrieved were cited less than 2% of the time (2025 data).
- "Summarize with AI" memory buttons
- We do not add buttons that plant instructions in a visitor's assistant. Microsoft found companies doing it and deployed mitigations.
Each row, its grade and its sources are on the Method page, including the rows that point the other way.
Terms every client gets
- Flat fees. One flat fee per site. The care plan is one flat monthly fee. Nothing is billed by the hour.
- You own it from day one. The client owns the domain, the content and the code repository from the first day, not at the end of a contract.
- Two revision rounds, then month to month. Every build includes two rounds of revisions. The care plan runs month to month, and if you leave, the site stays yours.
- Sign-off and accessibility. Every law-firm build ends with a lawyer's sign-off. Washington's advertising rule (RPC 7.1) governs what a law firm's website may say, and the lawyer stays responsible for it. Every site is checked against WCAG 2.1 level AA and evaluated by the practice itself. It is our own evaluation, not a third-party certificate, and we make no legal-compliance claim. Read the accessibility statement.
- No guarantees. We do not promise rankings, mentions or leads. We report the share of answers that name a business, with an interval, before and after.
We test on ourselves first
This site is build number one. It is made with the same kit and passes the same checks as every client site, and the assistants are measured on it the same way they are measured on a client. We froze the baseline design on 2026-10-11, before any real run: a fixed set of questions, asked many times on the official APIs, with our expectation written down before the data. The baseline will be published whatever it shows, including a result of no mentions at all. Read the research.
Founder
Still Fox was founded in Seattle in 2026 and is run by Josh Hsu. He tests a control before relying on it, asks how large a sample must be before a number means anything, and keeps a record someone else can check. The measurement engine follows those habits: a control cluster of prompts, intervals reported with every share, every answer stored. He reads every measurement before it is published.
Get a reading of what the assistants say about your business.
The snapshot is free and arrives by email in two business days.