Writing · SEO & Content · Dec 9, 2025 · 6 min

Running notes on GEO, AI search, and what actually matters

An hour-long GEO video I mostly believed, so I've been verifying it one claim at a time against my own site: I make the change, watch Bing Webmaster Tools, and re-ask three saved queries in ChatGPT and Perplexity every month. This page is the log.

4 minto register with Bing
~30 pagesindexed within a week

These notes started with an hour-long GEO video I watched in December and mostly believed. Belief is cheap, so since then I've been running its claims against my own site, one change at a time: make the change, then watch Bing Webmaster Tools and three saved queries I re-ask ChatGPT and Perplexity every month. Dated updates at the bottom; I rewrite sections when reality disagrees with them.

SEO grew a layer

The stat everyone opens with is that roughly 80% of searches still run through the classic Google box. Honest label: that's a borrowed number and I haven't verified it. What I can verify is my own analytics, since December, AI referrals to hi-daniel.com are a rounding error next to organic Google. Single visits on days organic brings dozens. The layer is real; the floor under it hasn't moved. Everything below assumes the page can already earn a ranking, because a page that can't rank doesn't get cited either.

The Bing dependency nobody prices in

ChatGPT's real-time browsing runs through Bing, mostly; ChatGPT itself hedges about post-GPT-4 behavior when you ask it. Either way, a site that never bothered with Bing hasn't been penalized by ChatGPT. It has excluded itself.

This claim was cheap to test, so I tested it. December 12: registered hi-daniel.com in Bing Webmaster Tools and submitted the sitemap. Four minutes against the advertised five. Within a week the crawl report showed Bingbot visiting daily and 30-odd pages indexed, pages Bing had no reason to know about before. Has a ChatGPT answer cited me since? No. Not on any of the three queries I track (“how to reconcile marketing attribution across ad platforms,” “best way to automate a weekly marketing report,” “what is GEO in SEO”). But in December those queries couldn't have surfaced me at all.

Evidenceone change at a time, on my own siteTested myselfran it, wrote down the resultPartly testedplausible, only my own recordBorrowedquoted, never verifiedGEO checklistThe version I pasted in Slacksix items, each with a test attachedBINGRegister with Bing4 minutes, ~30 pages indexedROBOTSUnblock the crawlersGPTBot, Bingbot, GooglebotEXTRACTAnswer in sentence onepaste the page into ChatGPTTABLESReal HTML tablespricing and comparisonsIDENTITYOne record, four views4 profiles checked, 2 driftedMENTIONSEarn the roundups7 of 8 tools, 2 roundup postsCitation scoreboardChatGPT 0 of 3 queriesPerplexity cited onceNext check: June
Every line carries a test and a result, which keeps the borrowed claims visibly separate from the ones I ran on my own domain.

Write for extraction

Every GEO post tells you AI reads structurally: question-shaped H2s, answer in the first sentence, tables for anything comparative. Fine, here's my one data point. In January I re-opened my Monday-report post, which began with two paragraphs of scene-setting before saying what the system does. I moved a one-sentence answer to the top and pushed the story underneath. That sentence says the reporting system I built reconciles nine data sources into one Monday number, automatically.

Then the cheapest extraction test I know: paste the page into ChatGPT and ask what it says. Before the edit, the summary led with my throat-clearing. After, it led with the sentence I actually want quoted. That's not a citation. I can't claim one yet, but it's the same extraction step, observed directly. An opener like “This depends on several factors…” fails this test every time; the extractor moves on to a page that doesn't defer. Answer first, qualify after. The prose can still be good. It just can't be the delivery mechanism for the fact.

Before · story first
Two paragraphs of scene-settingtop
ChatGPT summary led withthe setup
After · answer first
Nine sources, one Monday numbertop
ChatGPT summary led withthe answer
Same page, same question, two different summaries. The only thing I changed was which sentence went first.

The robots.txt check, in one paragraph

Confirm robots.txt isn't blocking GPTBot, Bingbot, or Googlebot. An accidental block removes AI visibility entirely and fails silently. I ran this on both my domains and wrote it up properly, missing subdomain file and all, in a separate post on the Ahrefs data. Score here: 2 domains checked, 0 accidental blocks, 1 subdomain with no robots.txt at all. Two curl commands. Go.

Machine-verifiable identity

The claim: AI systems cross-check who you are across public sources, and consistency is what earns trust. I can't see inside anyone's credibility model, so file this section under plausible, partially tested. What I could test was my own record. In February I pulled up every public description of me I could find. LinkedIn, this site, an old agency directory, a 2023 conference bio. Four sources; two had drifted. The directory still called me a PPC specialist, circa 2019. The conference bio named a company I'd left. I brought three of the four current; the directory ignores edit requests, which is its own lesson about controlling your record.

The honest limitation: I'm not in Wikidata or Crunchbase, and I'm not notable enough to force my way in. Parts of this checklist are only actionable for companies. For a company I'd treat LinkedIn, Google Business Profile, Crunchbase, and Wikidata as one record with four views, and reconcile it quarterly, because whatever the model actually weighs, “different answers to the same question” has never helped anyone's credibility, machine or human.

Third-party mentions are the citation currency

When a model answers a category question (best X for Y), it doesn't compose an opinion. It reads public ranking pages and splits the credit. This one is easy to observe yourself: I asked Perplexity for the best competitor ad intelligence tools and expanded the citations. Of the eight tools it named, seven traced back to the same two roundup posts. The model hadn't evaluated eight tools; it had read two listicles.

The two roundup posts behind Perplexity's answer. Both went on the outreach list.

So the model reads someone else's list. Your own page only files a claim. Find the roundups your category's questions resolve to and earn placement there. So PR now owns something checkable: get us into those two posts.

The version I actually pasted in Slack

What I sent a content team in March, verbatim minus the emoji:

  • robots.txt: confirm GPTBot, Bingbot, Googlebot aren't blocked. Subdomains too. Today.
  • Bing Webmaster Tools: register the site, submit the sitemap. Four minutes. Before lunch.
  • Every money page: question-shaped H2, answer in the first sentence. Test it, paste the page into ChatGPT, ask “what's the point of this page.” If the summary leads with your setup instead of your answer, rewrite the top.
  • Pricing and comparisons go in real HTML tables.
  • Ask Perplexity your five biggest category questions. Note which listicles it cites. That's the outreach list, send it to whoever owns PR.
  • Profiles (LinkedIn, GBP, Crunchbase, Wikidata if you qualify): one record, four views. Reconcile quarterly, calendar it.
GEO comes down to making your presence understandable, referenceable, and prominent enough that a model reaches for it.
the one-line version of these notes

The log

Dec 2025. First version, straight from the video plus the first tests: Bing registration done, robots.txt checked on both domains, the three tracked queries saved.

Jan 2026. Reformatted the Monday-report post answer-first; added the paste-it-into-ChatGPT test to the checklist after it caught the problem the rewrite fixed.

Mar 2026. Split the robots.txt audit and the Ahrefs correlation data into their own post. Rewrote the identity section after reconciling my own profiles (3 of 4 current). Citation scoreboard: ChatGPT 0 for 3 on the tracked queries; Perplexity cited this site once, on the weekly-report question. Small number. Real one, though. Next check: June.

Suggested posts