My AI search test log

What I am testing

From my post of 16 Sep 2026: "There is a version of your company that exists only inside ChatGPT. It was assembled from whatever you left in public, and it answers questions about you when you are not in the room."

I write about this on LinkedIn, so I test it on my own site, in the open. The site was rebuilt around three topics, each with its own page: AI content systems, content to drive B2B pipeline and LinkedIn programs for teams. Of the 25 test questions without my name, 16 are answered by a section of those pages, and 1 more by a page I added on 7 October (section 7). The other 8 are asked to see whether my name comes up where no page of mine answers. All of them were asked before the pages existed, and are asked again after Google has them.

How the test works

The test asks 27 questions, 3 times each, to two engines:

• OpenAI with web search: OpenAI's model with web search, asked through OpenAI's API. It is the closest stand-in I have for ChatGPT. The ChatGPT app itself can answer differently.
• Google's AI Overviews: the AI answer at the top of Google's results page, read as a searcher in the US, in English.

For each answer I count three things: does it name me (Mai-Lan Khong, in any of its spellings), does it cite a page of this site, does it cite my LinkedIn profile or one of my posts. A citation, in my own words from 9 Sep 2026: "A citation is the moment an AI answer links your post as its source. It is not a mention of your name and it is not a view."

Frontal, the company I work with, is counted apart and never as a point for me. When Google shows no AI answer for a question, that answer is left out, not counted as a miss. The main number is the 25 questions without my name. The two questions that carry a name in them are reported on their own.

The same question can get a different answer each time it is asked. In the first run, one of Google's three answers to one question cited my LinkedIn and the other two did not. So my rule, set before any run after it: one answer out of three moving on one question is not counted as a change. A change is read on the counts: for each engine and each of the three things I count, the share of answers received, on the 25 questions without my name together, and on each named question on its own. A count has changed when it is above its 28 September share in two full runs in a row, or below it in both. One run alone never shows a change, and a change first seen in December stays unconfirmed until a later run. Runs are compared only when they use the same questions, engines, model versions, number of answers, and Google's country and language. If one of those changes, a new series starts, and this page says so and shows no trend across it.

What I ask

27 questions, in five groups. I publish them word for word after the December run, not before: a page with my name next to each exact question could get me named for them, and then the test would partly be measuring this page. Keeping them off makes that less likely, it does not rule it out. This page is one of the changes (section 7), and any answer that cites it is reported on its own.

• AI content systems (9 questions, 6 answered on that page): among them, building a content system with AI for a founder, keeping the founder's voice, the documents a content team needs, and scaling to several founders.
• Content to drive B2B pipeline (6 questions, 5 answered on that page): among them, whether content brings pipeline and how to measure it, attribution, being cited in AI answers, and being ready for AI search.
• LinkedIn programs for teams (8 questions, 5 answered on that page and 1 on the page of 7 October): among them, how often to post, getting executives to post, engagement, content pillars, and whether personal content brings revenue.
• Who to hire or follow (2 questions, with no page of their own).
• With a name in the question (2 questions, reported apart): "Who is Mai-Lan Khong?" and one question about content engineering at Frontal.

The themes of the other 6 questions no page answers wait for December, with the questions themselves.

21 of them come from an earlier test of 7 Sep 2026. The other 6 were added on 28 Sep, kept close to words used by buyers and by people who sell to them.

Before: 28 September 2026

The first run was on the evening of 28 September. It came before the new homepage, About page and topic pages, but not before everything: the first technical fixes to the site went live one minute before it began. Google's copy of my homepage was then still from 21 September. So it is the before picture for the changes of 29 and 30 September, and nearly, not provably, for the fixes of 28 September. All 162 answers came back, and every Google question had an AI answer.

The 25 questions without my name (75 answers per engine)

The one exception: asked how to write LinkedIn posts that get cited by AI answers, one of Google's three answers cited my post of 9 Sep 2026 on exactly that. By my rule, one answer of three is not a change, so it stays a curiosity until it holds.

The two questions with a name in them (3 answers per engine)

Two of the three answers from OpenAI to "Who is Mai-Lan Khong?" cited my LinkedIn company page, which the count above does not include. It had no description then, and its type said "Self-Employed", so it was one of the first things I rewrote.

First look: 5 October 2026

The second full run was on Monday 5 October, less than five days after the three topic pages went live. Google had all three in its index, but had last read them on 30 September, so for Google the run came before the small corrections I made to them on 2 October (section 7). 161 of the 162 answers came back: one of Google's failed and is left out, not counted as a miss. Every Google answer that came back had an AI answer.

The 25 questions without my name (75 answers from OpenAI, 74 from Google)

Still no. The one Google answer of 28 September that cited my post did not come back. By my rule, one run alone never shows a change, up or down: the December run says whether these numbers hold.

The two questions with a name in them (3 answers per engine)

Against 28 September, each move here is one answer of three: OpenAI cited my LinkedIn once for "Who is Mai-Lan Khong?", and for the Frontal question Google cited this site and my LinkedIn once each, while OpenAI no longer cited this site for it. None of them is a change under my rule.

What I changed since

• 28 Sep: the site's technical fixes. One address per page, the text at the top of each page visible from the start, images from 29.5 MB down to 4.6 MB. The old meditation pages came down.
• 29 Sep: a new homepage and About page, with my one line on both. My LinkedIn company page rewritten with the same line, and my LinkedIn profile updated.
• 30 Sep: the three topic pages went live at 22:07. Google had all three in its index the same evening.
• 2 Oct: small corrections to the text of the three topic pages, and markup that names me as the person behind my two service pages. No title or heading changed. They went live three days before the run of 5 October, so an engine that read the pages again by then may have seen them.
• 7 Oct: a fourth page, on hiring for LinkedIn content: an agency, a content lead or your own team. It answers one of the test questions no page answered before.
• 22 Oct: a LinkedIn post from my company page on the same question as the hiring page: an agency or an in-house content lead.
• 26 Oct: small text fixes on five pages. My two service pages no longer cross out the options my hiring page calls workable, two pages no longer quote two of my LinkedIn posts, and my About page says since when I have been Content Lead at Frontal.
• 27 Oct: this page went live, listed here too because a page about the test is also something an engine can read.

Several changes landed in the same days, so if a number moves, this test cannot say which change did it.

What happens next

The full test ran again on Monday 5 October 2026, the first Monday after the three topic pages were in Google's index, less than five days after they went live. It runs once more on Monday 14 December 2026, about ten weeks after. Engines read new pages on their own schedule, so 5 October was a first look (section 6), and December gives them more time to read the pages. Neither run can show that a page caused a change. Each run is added here with its date, under the same rules, next to the one before.

Every Monday until 14 December I also ask one question, "Who is Mai-Lan Khong?", to see how the engines describe me. I read it each week and do not publish it each week: 3 answers per engine a week are too few to read on their own.

What this test cannot tell

It asks two engines. Perplexity, Claude and Copilot are not in it. A mention or a citation is not a visit, and a visit is not a client. Even the visit is hard to trace, because, from my post of 14 Sep 2026, "a visit from an AI answer carries no record of which post it quoted". And it is one site and one person: what happens here is what happened here, not a promise for anyone else's site.

Where to go next

• AI content systems: how the machine behind the content works, and how it keeps a human voice.
• Content to drive B2B pipeline: content for one buyer, the order of channels, and being found in AI answers.
• LinkedIn programs for teams: the founder, executives and employees posting together.