Two Free AI Scanners Graded Our Week-Old Site
We sell a paid AI-visibility audit. A growth agency gives two away free. We pointed both of its scanners at our own methodology page and published everything that came back, the fabrications and the two findings that were right.
How this page came to be
We sell a paid version of what these tools give away, so read this page knowing that. We ran two free AI-visibility scanners from the same growth-marketing agency against our own methodology page to see what a free scan actually tells a business about itself. We publish what came back, including the two findings that were correct about us.
- Prompted by
- Bryan Markham, CitedScore's founder
- Provided
- the public URL citedscore.com/how-it-works, entered into each scanner exactly as any visitor would enter it
- Conducted
- August 27, 2026
- Model
- Two free AI-visibility scanners from one growth-marketing agency. Neither report discloses which underlying models produced its text, so we name what is verifiable and no more.
- Limitations
- One page, one day, two tools from a single vendor: nothing here generalizes to every free scanner. We also cannot see either tool's internals; where we say a number looks like a fallback constant, that is an inference from the number itself, and it is stated as one.
- The receipt
- We are not naming the two scanners or the agency that operates them, and we are not linking the reports, because the links carry the vendor's name and this page is about a category, not a competitor. Every quoted line below was captured on August 27, 2026 from the two public report pages as they rendered that day (their score accordions are client-rendered and were expanded before extraction), and the full set of quoted lines is committed to our repository as the fixture our test suite checks this page against; the report URLs and re-capture notes are recorded in issue 175 of our public tracker. One exception is flagged where it appears: the battlecard's session-replay analytics row is described without a committed line, because quoting that row would name an unrelated real product.
The number, before the receipts: both reports put our average content age at roughly five years. The domain was registered on April 28, 2026. The site launched on August 19, 2026. The scan ran on August 27, 2026, eight days later.
What we ran
On August 27, 2026 we ran citedscore.com/how-it-works through two free AI-visibility scanners operated by the same growth-marketing agency. We are not naming the agency or either tool, for the reason the receipt note above states: this page is about what free scanners do, not about one company that makes them. Every line quoted below is transcribed verbatim from the two reports as they rendered that day, and the full set of quoted lines is committed to our repository as the fixture our test suite resolves this page against.
Three dates matter for what follows. The domain was registered on April 28, 2026. The site launched on August 19, 2026. The scans ran on August 27, 2026. The page the scanners read also states its own age twice: a visible line reading Method last reviewed 19 August 2026, and a machine-readable dateModified of 2026-08-19 in its WebPage schema. Most of what follows is arithmetic against those dates.
The five-year content age
The first scanner's freshness finding: “Your average content age is roughly 5 years. AI deprioritizes stale pages.” The second stated the same figure with more precision: “Your Grounding score is 75/100 and one of the two remaining gaps is content freshness: average page age is approximately 1,825 days.”
A site that has existed for eight days cannot carry five-year-old content. And 1,825 is 365 times 5, exactly. A measured average lands on an untidy number; a default lands on a round multiple of a year. We cannot see the tool's internals, so we state this as the inference it is: that number reads as a fallback constant narrated as a measurement. What we can state as fact is the arithmetic, and what the advice attached to it prescribed: “Audit, update statistics, add a published or reviewed date, and add at least one scannable list or table per page”. The page it was advising already carries a published-and-reviewed date twice, once in visible copy and once in machine-readable schema. The remediation prescribes a fix the page already ships.
The mechanism this contrasts with: every finding in a CitedScore Report carries an evidence tier, Measured in this audit or Standard practice, so a reader always knows whether a number was observed in their run or is an industry baseline. A value we could not measure reports as not measured. There is no constant to fall back to, because a constant dressed as a measurement is the exact failure this page is documenting.
A score that disagrees with itself
The second report's recommendations panel says: “Your Prominence score is 0/100 and your share of voice is 0% across all tracked answers.” The same report's scores panel, verbatim: “50 Prominence How often you surface in AI answers for your space.” Same metric, same page, same run, two values.
The 0 agrees with everything else in the report's own data, including the answer sample below. The 50 agrees with nothing, which is why it reads as a display default sitting where a measurement should be. Either way, the report contradicts itself on its own headline metric, and the contradiction shipped.
The mechanism this contrasts with: human review. Every CitedScore Report is reviewed by a person before it reaches a client, and a score panel disagreeing with its own recommendations is the first thing a human reader catches. A pipeline that ships its output unread will eventually ship a contradiction, because nothing in it is looking.
Competitors that do not exist
One report includes a competitor battlecard. Three of its rows, verbatim: “Growthract 1 3%”, “Prometiv 1 3%”, and “Synthic 1 3%”. One appearance and a three percent share each, and we can find no company operating under any of those names. We print them here precisely because there is nothing to find: naming a real competitor would be an accusation, and naming these is a demonstration. Alongside them, the battlecard listed a session-replay analytics product, categorized as an AI-visibility auditor; that row is the one line on this page described without a committed fixture quote, because quoting it would name an unrelated real product.
This is what unvalidated model output looks like when it is transcribed straight into a deliverable. Language models produce plausible brand names on request; that is not a flaw, it is what generation is. The flaw is publishing the output without checking whether the companies exist.
The mechanism this contrasts with: brand validation. A competitor name that surfaces in our pipeline has to resolve to a real company before it can appear in a Report; a name that cannot be validated is dropped, not printed. The battlecard above is the counterfactual: what a competitor section looks like when that step is missing.
A sentiment pass on zero mentions
One report graded us a pass on “Positive brand sentiment” and a pass on “Third-party sources vouch for you”. The same report's own answer sample: “You show up in 0 of 30 AI answers”.
Sentiment is a property of statements. Zero statements have no sentiment. A brand that appears in none of the sampled answers cannot have positive sentiment in them, and third-party sources that never mention you cannot be vouching. A pass here is not a generous reading of thin data; it is a grade assigned to data that does not exist.
The mechanism this contrasts with: a metric with nothing to measure reports no data, and the Report's vocabulary rules ban the hedge words that let an unmeasured value pose as a finding. Words like might and we believe are forbidden in Report copy outright, because a sentence that needs them is a sentence without a measurement behind it.
A proxy as a publisher to pitch
One report recommends publishers to pitch for coverage. Its third row, verbatim: “vertexaisearch.cloud.google.com third-party article 5x Not in it”. That hostname is not a third-party article and not a publication. It is the redirect proxy Google Gemini's grounding citations pass through on their way to the real source. There is no editor there to pitch; a pitch sent to that suggestion would be a pitch sent to a piece of Google's plumbing.
The detail is worth pausing on, because it cuts both ways. A proxy hostname appearing in that list means the tool is reading real citation logs from real AI answers, and passing the hostnames straight through without asking what they are. The data underneath is genuine. The reading of it is not.
The mechanism this contrasts with: CitedScore records citations only from each platform's structured source list, the machine-readable list of sources the platform itself attaches to an answer, and classifies what it finds there before any of it reaches a Report. Infrastructure hostnames are infrastructure, not pitch targets.
Demand numbers from nowhere
One report states there are “8,400+ monthly AI-driven searches” in our topic, with “+78% demand growth, YoY”.
No AI platform publishes per-topic answer volumes. There is no public dataset those two numbers could have come from, and no year-over-year series for a metric nothing public measures. A number with no possible source is not an estimate; it is set dressing.
The mechanism this contrasts with: every third-party statistic on this site resolves to a named study at /sources, with the publisher, the date, and the figure as published. A number we cannot source does not appear. That rule costs us impressive-sounding paragraphs regularly, and it is not negotiable, because this page is what the alternative looks like.
What was real
Two findings survive the receipts, and they deserve to be stated as plainly as the fabrications. First: “You show up in 0 of 30 AI answers”. Second: “You are in none of the sources AI cites. Rivals are named 58 times to your 0.” Both are true. Our own Report runs in early August drew the same picture from a larger sample, and we published those numbers in the first post in this series rather than waiting for flattering ones.
A brand can also be cited without ever being named, which is its own failure mode: the ghost citation. That term is not ours. It comes from Semrush and Kevin Indig's June 2026 Ghost Citations Study, which named the problem and measured it; the full citation is at /sources. We measure where a brand falls on that spectrum per audit. We did not coin the vocabulary, and this page should not leave you thinking otherwise.
So the free scanners and the paid audit agree on the one thing the free scanners actually measured: eight days after launch, AI answers did not know we existed. The difference is everything wrapped around that finding.
What a free scan buys
A free scanner has one real data source and a template with many slots. Where the data reaches, you get a real finding, and the absence finding above is one. Where it does not, the slots still render: a constant becomes a content age, a display default becomes a score, generated names become a competitive landscape, an empty sample becomes positive sentiment, and a proxy hostname becomes a publisher to pitch. Nothing in the report tells you which kind of line you are reading.
That distinction is the product we sell, stated once and without a discount: free scanners guess where the data runs out, and a CitedScore Report labels every finding Measured in this audit or Standard practice and declines to invent the rest. Every mechanic named on this page, tiers, validation, source lists, review, is documented in full at /how-it-works, on the same page the scanners graded.