Who is behind LLMention?
Alex Xu builds and maintains the project: the scanner, the rule catalogue, the study and this writing. The byline is a handle rather than a legal name, and the way to verify it is the public repository at github.com/Alex13192/geo-scanner, where the rule set and the site itself are readable.
That is a deliberate trade. A pseudonym with a public repository and a working contact address can be checked; a full name with no work attached to it cannot. If a legal name is required by a payment processor or a regulator, it is provided through the operator details on the refund policy rather than invented for a byline.
What qualifies him to run this measurement?
Nothing in the way the question usually means. There is no data-science credential and no SEO agency history to cite, so the page does not claim one.
The qualification offered instead is structural: the method is published in full, every rule and its weight are on the methodology page, and the scanner runs the same catalogue the page publishes. A reader who thinks a rule is wrong can name it and check it, which is a better guarantee than a biography — and the only one this project can honestly make.
What is his professional background?
Alex Xu has spent more than two decades in network engineering, most of it on large enterprise and carrier networks where a crawler or a misconfigured edge is not an abstraction. He is a Cisco CCIE (Routing & Switching and Security), CCNP and CCNA, Microsoft MCSE and Juniper JNCIS certified.
His work has spanned data-centre and WAN architecture, hybrid-cloud and SD-WAN transformations, network security, and automation for very large, distributed organisations. Specific employers, dates and scale figures are not published here: this site is run by an independent developer, and the professional record is available on request.
The specialities that shape this site’s checks are the same ones: data-centre fabrics (VXLAN/EVPN, spine-leaf), BGP/MPLS/SRv6, multicast and low-latency switching for AI clusters, network security (Zscaler, Palo Alto, FortiGate, Cisco ASA/Firepower), and automation with Python, Ansible and Nornir. Crawler access, edge delivery and machine-readable structure are the parts of GEO that are network problems, and they are written by someone who has spent two decades on network problems.
A full résumé, including employers and education, is available on request at services@llmention-geo.com. Phone number, postal address and date of birth are deliberately not published here.
How is the writing produced?
Every piece starts from a measurement, then goes through the same four stages. The drafting stage uses a language model; the three checks around it are what make the result publishable.
| Stage | What happens | What would stop publication |
|---|---|---|
| 1. Data | The study rows or the rule catalogue are read, not remembered | A figure that does not appear in either |
| 2. Draft | A language model prepares the first version against that data | Drafting happens before the data, which is how invented numbers get in |
| 3. Check | Every number, name and status code is traced back to its source | One untraceable figure |
| 4. Publish | The limits of the claim are written into the piece, not left out | A headline the data does not support |
Saying the drafting stage uses a model is uncomfortable and true. The alternative — claiming every sentence came from a person — would be a false statement on a site whose entire argument is that claims should be checkable. The standard is not who typed it; it is whether each number can be re-run and each limit is stated.
What are the editorial rules?
Five rules, all of them checkable against what is published, and all of them broken somewhere by sites in this category.
- No sponsored content and no paid placement. A report is a report; a mention is not for sale, at any price.
- No number without a source. Percentages, status codes and weights are traced to the study or the catalogue before publication.
- Limits are published with the claim. The scan does not ask a model about your brand, and every page that implies otherwise is being corrected.
- Corrections go where the error was. A wrong figure is fixed on the page that carried it, with the date changed.
- No volume targets. A piece is published when there is a measurement behind it. There are six; there is no quota of thirty.
The last rule is the one that makes the archive look thin on purpose. Of the 30 homepages in the study, 53% graded F, 30% never answered the crawler at all, and none reached an A — the findings are worth writing up, but a blog that padded them to a weekly schedule would be imitating the content this project measures against.
The weights that decide those grades are published, and they are the reason the writing goes where it does: AI crawler access carries 16%, machine readability 12% and content depth 11%, while the context file this site still publishes is capped at 3.3 of the 100. A claim about the small lever is not worth an article.
How is the site funded?
Two ways, both printed on the site rather than discovered later: advertising served by Google, and the paid measurement reports described on the pricing page, from $39. The prices are imported from lib/pricing.ts and are the same numbers the checkout uses; when they change, they change in one place.
The obvious conflict is that a paying customer might prefer a better number. Three things constrain it: the scan is free and identical for everyone, the score is computed by code from the response it fetches, and the paid report explains a result rather than changing it. A reader can verify the first of those by running a scan, and the other two by reading the source.
There are no affiliate links, no paid directory placements that carry a rating, and no vendor logos that imply an endorsement. Where an outbound link earns anything, it does not appear on this site.
How do I report a mistake?
Write to services@llmention-geo.com with the URL, the check and what the scan reported. The most useful message this project receives is a disagreement with a specific rule, because the rule set is public and a specific objection can be answered or acted on.
Corrections to the writing follow the same route. A figure that turns out to be wrong is the one thing this archive cannot quietly fix, because it is the thing the rest of it rests on.
What does the site deliberately not do?
It does not ask an AI engine whether your brand is mentioned, and it does not pretend a readiness score is a citation. Those limits are stated on the About page and in every article that touches the question.
It also does not publish a method it cannot execute: the tests assert that the published rule set and the scanner’s behaviour cannot drift apart, which is a stronger promise than a methodology page written by hand and never checked again.