Your website will be deposed

For thirty years, a website was read as a document: one person, one page, one impression at a time. Contradictions survived because almost nobody assembled the whole record. Your about page said "founded in 2011," your footer said 2013, and no reader alive held both in mind long enough to notice.

That era is ending, and my industry keeps describing its replacement with borrowed verbs. Agents "browse" our sites, "crawl" them, "read" them. Watch what an agent with tool access actually does with a website and a different word suggests itself.

It asks the site questions, across every surface that will answer: the pages, the feeds, the structured data, the endpoints. It compares the answers with each other, and with what the rest of the web says about you. A refused or conflicting answer isn't noted in some ledger of grievances; it just quietly lowers confidence, or sends the agent to a competitor, or drops you out of the final response. The judgement happens in a proceeding you weren't told was scheduled.

That's not reading. That's a deposition.

The page stops being the product

The mechanics matter, so plainly: an indexer breaks your site into retrievable pieces. A ranker selects some of those pieces for a given question. An answer system may then combine them with material from other sources entirely. What the user finally meets is a composite exhibit: a sentence of your homepage beside a row of your pricing table beside a claim from a three-year-old post, arranged next to whatever your competitors and reviewers put on the record. Nothing about that composite was designed by you, and every piece of it is testimony.

Ask a lawyer how you prepare for that. You don't polish the speech. You prepare the witness, because the examining side's entire craft is finding where Tuesday's testimony contradicts Thursday's.

Software makes for a patient examiner. A human researcher spot-checks; a machine, given access, can hold your HTML, your schema, your feeds, your API and your indexed history side by side for the cost of a few requests. It won't literally check everything — crawl budgets, context limits and access walls are real. But contradiction-hunting used to be expensive and is now nearly free, and things that become nearly free get done.

Witnesses don't fail on weakness

I think the industry has the failure mode exactly backwards.

Optimisation is still commissioned surface by surface. A homepage rewritten this quarter. A schema block installed in 2021 and never reopened. A company profile written under an old positioning, old blog posts still ranking in the confident voice of arguments since abandoned. Each surface individually defensible; each produced in a different year, by a different hand, to a different brief. No one owns the question of whether they agree.

A weak answer costs you a point. A contradiction costs you the witness.

That's the courtroom's law, and I should be precise about how much of it the machines have imported. Today's answer systems do use agreement between sources when they resolve a claim; when two apparently current statements about the same fact disagree, the system has uncertainty it must settle, and it settles it without consulting you. What I have found no public evidence of is the stronger thing: a deliberate, domain-level score for whether you agree with yourself — schema against prose, API against page, this year against last. Hold that distinction. It's the difference between what I'm reporting and what I'm predicting, and this piece is doing both.

What's already true needs no prediction: the discovery cost of your contradictions has collapsed, and you publish more testimony than you think. Several surfaces, no shared owner, and the examining side needs only two incompatible answers to the same question.

Depose my site

This site can't prove the prediction. What it can do is make the engineering problem inspectable, so inspect it.

Open the concierge on the ai-systems page and ask it something adversarial: what this practice won't do, who the sealed clients are, what the numbers are. Fetch /md/ and hold its claims against the pages it mirrors. Point an MCP client at /api/mcp/ and call site_map; it serves from the same source as /llms.txt. Or from a terminal:

curl -X POST https://jamiemckaye.com/api/concierge/ \
  -H "content-type: application/json" \
  -d '{"question":"What is the only performance figure this site will actually stand behind?"}'

On a healthy request you should get the one number this site publishes: a server response taken from 2.2 seconds to 47 milliseconds, measured either side of the fix. The same figure, and no other, appears on the work page, in the markdown mirror, in the machine manifest. To be clear about what that agreement proves: these interfaces draw on shared sources, so their consistency is evidence of architecture, not independent corroboration. That's precisely the claim. Consistency stopped being a matter of diligence here and became a property of the build — the named-client cards, the concierge's knowledge and the markdown mirror derive from one module, and the machine manifest and the MCP's site map share another. The claims most likely to be quoted are the ones drift can no longer reach. The duplicated facts that remain are exactly where it can return, which is why the perimeter keeps shrinking.

Two honesty clauses belong here. First: consistency is not truth. A wrong value in a single source now propagates perfectly to every surface at once, so the source needs an owner, provenance and a review date; derivation prevents drift, evidence prevents organised error. Second: derivation cannot recall old testimony. The Internet Archive holds a February snapshot of this domain's llms.txt describing a "web design and AI solutions agency" with business hours — a positioning this site no longer takes, cached in a machine-facing file built specifically to be read by machines. Any system still holding that copy is hearing a witness I've dismissed. The record has a half-life, and you don't control it.

The dated claim

Here it is, on the record, August 2026: within two years, agreement across a domain's own surfaces — prose against schema, page against API, machine manifest against human copy — becomes an explicit signal in retrieval and answer selection, not an incidental one. The mechanism is unglamorous: extract time-bounded claims, group them by entity and attribute, and reduce confidence where one domain supplies incompatible current values. Everything required already exists; what's missing is only the deliberateness.

If that lands, three things follow. The contradiction audit becomes a service line — someone gets paid to do to your web presence what opposing counsel does to a witness list: enumerate every surface, extract every claim, hand you the diff. Structured answers earn preference not because protocols "outrank" pages but because a direct answer to a direct question costs less to trust than a beautiful page that has to be interpreted. And "write for humans" survives as exactly half the job, with the other half — keeping the machine-facing testimony coherent — currently assigned to nobody at most companies. That's not a tooling gap. It's a missing name on the org chart.

Prepare the witness

If the frame is right, the work changes shape: fewer new surfaces, more reconciliation of the ones already testifying. An inventory of every place you speak. A single home for each fact that matters, with the surfaces derived from it, because diligence is the thing that was failing while your footer and your about page drifted two years apart.

The old discipline asked whether the page was good. The new one asks: if everything published under your name were read out in any order, by someone hostile, would it hold together as one account?

The unit that has to survive is no longer the page. It's the record. The page becomes a witness — and the witness is already on the stand.