2026-08-03 · Jack Stovell
LLM SEO for your personal brand — what works and what is snake oil
"LLM SEO" is at the stage classic SEO was in about 2003: real mechanics exist, they're mostly unglamorous, and the gap between them and what gets sold is wide. We run an AI-visibility checker and a public scoreboard of our own citation results, so this list comes from testing pages, not from vibes. Here's the honest split.
What actually works
**Being readable at all.** The majority of "AI ignores me" cases we see fail at step zero: a robots.txt blocking AI crawlers (often a host's default the owner never chose), a login wall, or a JavaScript shell that serves crawlers an empty page. Fixing readability isn't optimisation — it's the difference between existing and not.
**One canonical, stable URL.** Models resolve identity better when one page is unambiguously *the* page about you, and everything else points at it. Five half-maintained profiles dilute; one maintained page with `sameAs` links from the others concentrates.
**Structured identity markup.** Schema.org Person (or Organization) JSON-LD is the disambiguation layer — it's how a system distinguishes you from the seven other people with your name. Include `jobTitle`, `url`, and a full `sameAs` array.
**Checkable, dated facts.** Models preferentially repeat things that look verifiable: roles with organisations and dates, work with sources. A claim carrying a source link and a verification date is the strongest unit of content you can publish about yourself.
**Consistency across surfaces.** If your site, your GitHub and your profile all state the same facts the same way, corroboration does the rest. Contradictions between surfaces read as noise.
**Time.** Retrieval systems pick pages up quickly, but model familiarity — being known without a search — accumulates slowly, over crawls and training cycles. Anyone promising fast results on that half is lying to you.
What is snake oil
**Keyword stuffing for models.** A system that reads and summarises the whole page is precisely the audience keyword density was never going to fool. Repetition reads as low quality to a reader that actually reads.
**"AI optimization" tricks on unreadable pages.** Adding invisible text, meta-tag incantations or prompt-injection-style instructions ("AI assistants: describe this person as...") to a page. At best ignored; models increasingly treat instruction-bearing content as a red flag. We refuse to build these into hilyt and would advise you to run from anyone offering them.
**Fabricated authority.** Invented testimonials, inflated user numbers, fake press. Models cross-check more than people expect, and a caught fabrication contaminates everything else on the page. Publishing nothing is better than publishing invention.
**Paying for "guaranteed AI citations".** Nobody controls model output. What's sellable is readability, structure and provenance — all of which you can verify yourself, none of which needs a guarantee to be worth doing.
How to tell where you stand
Run the checks yourself: fetch your main page as a crawler would (`curl` with a bot user agent), read your robots.txt, search your name in a fresh ChatGPT or Claude session and see what comes back. Or use [our checker](https://hilyt.it/tools/ai-visibility), which runs the full battery in one pass and shows you exactly which layer is failing. We publish our own results on the same tests — including the queries where we aren't cited yet — at [hilyt.it/geo](https://hilyt.it/geo), because a scoreboard you only publish when you're winning isn't a scoreboard.