Accuracy

Accuracy answers one question: is AI telling the truth about you? Trakkr asks ChatGPT, Claude, and Gemini the questions your buyers ask, with live search switched off, then grades every answer against your own site, the facts you confirm, and an independently retrieved record of who you are.

Live search is off on purpose: what a model says from memory is what a buyer gets when they ask once and take the answer at face value, which is where a wrong answer sits unnoticed for months. Citations covers the search-first surfaces. Open Accuracy when a prospect quotes a price you do not charge, or when answers about you keep naming someone else.

The first scan

A brand that has never been checked gets a first-check screen. It shows an example of a wrong answer caught in a check, names the AIs it asks, and says what it asks about: buyer questions Trakkr drafts from what your company does (six by default, each staged as Discovery, Evaluation, or Diligence), plus who runs, owns, and founded the company. Select Start the first check. The scan runs in the background and survives a reload, and the questions fill in as they are checked. Weekly checks follow. Viewers see the screen with no button.

The masthead

The report opens on the Buyer questions tab with two readings side by side.

Wrong answers counts the answers that are false at the latest check. It shows the change since the check before, a line over the last 12 checks (each wrong answer put right is marked in green), the date of the next check, and how many answers each AI gets wrong now. A wrong answer only counts once the error repeats in 2 of 3 asks. Select the number to see the wrong answers grouped by question.

Needs your answer counts claims an AI made that nothing on file can settle. Each one takes a tap: True, False (then say what is true instead), or Doesn't matter. Your answer is used from the next check.

A check that could not read everything is never shown as clean. When an AI or an evidence file did not come back, the Wrong answers reading says some answers did not come back, and a notice under the tabs says the check has partial coverage. Missing data is never counted as correct.

The question matrix

One row per buyer question, one column per AI, then the question's Harm, a 12 weeks line, and a Since date for when its answer last changed. Rows are sorted worst first. A cell is that AI's verdict on that question at the last check.

Select a row to open the question. The view pills above the table narrow the list to All, False, Sends elsewhere, Unsettled (nothing false, but not all checking out), or Checks out. Search and the Filter menu narrow it by wording, how often a question is checked, or AI. With the list focused, j and k walk it, Space peeks at a row, and Enter opens it.

The list is yours. Select Add a question to measure a new one straight away, up to 25 per brand. A row's menu lets you edit its wording (past checks keep the old wording), check it weekly or monthly, or archive it. A question right eight checks in a row drops to monthly, and any false answer promotes it back to weekly.

Reading one question

Opening a row gives you the question with each AI's answer under What each AI said, and the grading as annotation.

• The truth sits at the top: what is true, with its source and a link to the Verified record.
• Each answer is quoted word for word, exactly as that AI returned it, with search off. Wrong parts are struck through and numbered.
• Each number maps to what is true instead, with its source: "You confirmed", your site, or the verified record.
• An unsettled claim carries an Is this right? prompt inline, and your answer saves to the verified record.

The foot offers Plan the fix on a false answer (it puts the fix on your action plan). When your page is already right it offers Add monitoring action. Ask Agent and Check history are links in the drawer head. Re-check this question and Archive question are in the head's menu, or in the foot when there is nothing to fix.

Where the truth comes from

Identity checks cover the CEO, public or private status, founders, parent company and headquarters. Trakkr compares assistant claims with reference facts supported by agreeing search evidence. Read the receipt and its source before treating a verdict as settled: agreement supports confidence but is not proof that a fact is current. Confirm or correct an unsettled claim through Is this right?.

That scope is deliberate: independent search suits identity facts and is weaker on plan names, prices, trial terms, integrations, certifications, and who a product is for. The site pass handles those. Your website is the authority for what you currently publish, so Trakkr reads the pages most likely to answer each buying question and compares the assistant claims with the page text. Prices, certifications, and company numbers are pulled out with deterministic rules and prices compared with arithmetic, keeping a model's memory out of the evidence.

The four site rows

The audit file lists every claim with a status. Four of them describe how your own site compares.

Unpublished is the most useful row and the easiest to misread. It does not say the assistant is wrong; it says buyers are getting an answer your site does not support, and assistants are filling the silence from somewhere else.

Price disagreements show both sides on purpose. Calling $18 per user per month wrong against a page that says $20 per member per month would mean guessing the billing period, the region, and whether a user and a member are the same unit. Stating both sides stays true even when an annual discount later explains the gap.

The fix block

A proposed change appears as a diff: the line to remove, the line to add, and the page it belongs on. The removed line is never written by a model. It has to be an exact substring of an unexpired copy of the page saved for that check, and a near miss means the page changed underneath us, so nothing ships and the next crawl tries again. An edit is only offered when that line contradicts a fact you confirmed.

An Unpublished row gets a draft instead: nothing is removed, it uses only confirmed facts, and nothing is published for you. A Site differs row never gets an edit, because that row declares no winner. Site findings that need an owner also appear as suggested work under the questions, ready for your action plan.

Corrections and check history

The Corrections tab lists what stopped being wrong: the question, the AI, what it says now, a track of its checks, how many checks it was wrong for, and the day it was put right. Newest comes first, and you can narrow the list to one AI. Only landed corrections appear; one that goes wrong again reopens.

The Check history tab has two columns. On the left is Every check, newest first, with what changed since the one before and the false count at each. On the right is the check you opened: its verdict, What changed since the previous check (fixed, new, or regressed), Every answer in this check, and How the check ran. A partial check is marked as not counted so it cannot flatter the trend. Two links at the top of the check open files:

• Buyer conversations holds the recorded multi-turn threads, replayed as the buyer saw them.
• The audit file is the whole check: how it ran, what each AI said, which of your pages were read, and every claim with what happened to it.

The Verified record tab groups every fact by who settles it: verified independently, only your site can settle, or you confirmed.

Re-check, export, and print

To check the whole report again, open More ways to check beside Add a question and select Re-check everything. It asks every active question again, reads your pages, and grades every answer, in the background. Manual full checks are capped at one every seven days. While the wait is open the dialog says when the next full check is and points you to a single question instead, which has no weekly limit: select Re-check this question in its drawer. You can also refresh the facts from live search for a full check, which is slower.

Select Export record, choose a format, then confirm. Printable record opens a print-ready page at /accuracy/print in a new tab, and you save it as a PDF from there. White-label is the default inside a client portal and an opt-in elsewhere (Hide Trakkr branding). Claims as a spreadsheet downloads a CSV, and A short summary copies plain text for an email or Slack. A partial check cannot be exported: the dialog offers the last complete check instead.

Limits Trakkr states out loud

• Every quote attributed to your site must be verbatim from a page Trakkr fetched. If it is not in the retrieved text, the row drops to Inconclusive.
• A dated page (a press release or news post) can show something happened. It can never establish a current CEO, price, owner, or offer.
• Unpublished is only claimed for a question whose answering page Trakkr read. Anything outside the areas it covered stays unchecked, not missing.
• If your domain blocks every read attempt, the report says so and grades identity claims against independent sources only.
• A scan missing an AI or an evidence file is saved as a partial check. Missing data is never read as correct.

Common questions