Lots of people are talking about GEO (getting AI to recommend you) these days, but almost no one spells out the more practical question: after you do it, how do you know whether it actually worked?
We ran a serious test round on this ourselves. In early August, we put our own company in front of AI models and asked several types of questions. The most painful one: we asked “Which reliable IT operations outsourcing providers for small and medium-sized businesses are there in Beijing?” The AI recommended 5 companies — not one of them was us.
After that we did a few targeted things (we’ll cover which ones below), and five days later we re-tested with the same type of question. This time, on one of the platforms, we appeared first in the “top recommendation” category, along with a fairly accurate description of our business.
But let me be clear up front: five days, one re-test, an extremely small sample — this cannot prove causation — AI answers are inherently random, and changing a single word in the question can change the result. So the point of this article is not “we succeeded,” but rather the four things we found during testing that are more useful than whether we got a hit or not.
Finding 1: Results vary enormously across AI platforms — testing only one is bound to mislead you
We tested three platforms at the same time, with the same questions, on the same day. The differences were far larger than we expected:
| Question type | Platform A | Platform B | Platform C |
|---|---|---|---|
| Brand term “What does XX company do” |
? Accurate | ? Accurate | ? Accurate(but omitted qualifications) |
| Product term “What is XX product” |
? Detailed | ?? Mentioned only | — |
| Category term “Which XX providers are there in Beijing” |
? First in top recommendations | ? Not present at all | ? Not present at all |
Same day, same questions: one platform ranked us first in its recommendations, while the other two didn’t mention us at all. If I had only tested the lucky one that day, I would have concluded “our GEO is working well”; if I had only tested the other two, the conclusion would have been “it has no effect whatsoever.” Both conclusions would have been wrong.
So the first methodological conclusion is firm: test at least three platforms, and fix it so that you test the same ones every time. Different AIs draw on wildly different content sources, so results from a single platform are not representative.
Finding 2: Don’t just look at the answer — look at what it cites (the most useful one)
This was the biggest takeaway this time. Most people testing GEO look at only one thing: am I in the answer. But what can actually guide your next move is the list of sources it cites.
When answering the same type of question, the three platforms cited completely different things:
- One platform pulled from B2B directory sites. It got our business right, but the basis was not our official website — it was some corporate information directory page, and that page’s content was years out of date.
- One platform pulled from our official website — the “About Us” and “Solutions” pages — where the information is the most current and accurate.
- One platform pulled from media coverage. When it recommended us, the reference sources listed at the end were reports from several financial and industry media outlets.
The value of this information is that it tells you directly where to put your money and effort. If a platform mainly feeds on media coverage, then for it, ten articles on your own website may matter less than one industry analysis published by a legitimate media outlet; if a platform mainly feeds on your official website, then making your website information accurate and complete is the best-value move. There is no universal playbook — only “what does this platform feed on.”
Finding 3: AI will tell you what it’s missing — this is free diagnostics
When answering about our company, one platform added a line at the end: “The search results mainly provided business registration information and a business introduction; no specific qualification grade certificate information was found.”
That line looks like a disclaimer, but it is actually a free gap diagnosis. We do hold ISO management system certifications, software copyright registrations, and industry qualifications, and they are listed on our website — but within that platform’s retrieval scope, there is no source it trusts for any of this.
The reason isn’t hard to guess: for things like qualifications, stating them on your own website is “self-reporting,” which machines trust less; what it needs is a third-party source it can cross-check. That turns into a very concrete to-do — get your qualification information backed by retrievable third-party endorsements, rather than repeating it over and over on your own website.
So after testing, don’t just look at the hit rate. When AI says “I couldn’t find XX,” it is telling you directly what to fill in next. That is more direct than any SEO diagnostic tool.
Finding 4: To get onto a “recommendation list,” you first have to get into someone else’s “roundup article”
We specifically studied the two platforms that “didn’t mention us at all,” to see what they based their recommendations of others on.
The result was clear: what appeared repeatedly among the cited sources were roundup articles like “Which XX Companies Are There in Beijing,” along with “recommendation lists” and “reputation rankings” — one of those roundup articles was cited 4 times.
This explains a phenomenon many people find puzzling: no matter how well written your website is, it won’t help with questions like “recommend a few.” To assemble a list, AI needs list-style content that itself already names multiple companies; your website talks about only you from start to finish, so it has no way to “extract” a list from it.
Following this logic, there are just two directions for making progress on category terms: get yourself into roundups/lists that others produce, and produce genuinely objective comparison and selection content yourself (explaining what kinds of this service exist and who each suits, and placing yourself honestly within one category). Self-introduction alone won’t get you there.
What we actually did during those five days
To make the comparison above meaningful, we should spell out the actions too. Three things, all basic work on the website side:
- Put the company’s identity information into one machine-readable block: full name, short name, year founded, address, phone, core business and service area in one place, with structured markup for machines, not just layout for human readers.
- State the correct name explicitly and call out the common misspellings. We had found earlier that AI would write our company name with homophone typos, so we put a right-versus-wrong comparison directly on the page. If you don’t write it, the machine can only guess.
- Add comparison tables to articles about comparing options. The content was good, but it was all paragraphs, with no table that could be quoted as-is.
To stress it once more: you can’t claim these three things are responsible just because something changed five days later. The credible approach is to keep them fixed and record continuously for several months to watch the trend. GEO is a slow, month-level variable, and the noise in any single result is larger than the signal.
How to build your own GEO measurement: three layers × three platforms × once a month
- Layer one: brand terms: “What does Company X do?”, “How is Company X?” Check whether the facts are accurate. This layer is the easiest to pass. Don’t get cocky when it does pass, because anyone who types your full company name already knows you.
- Layer two: product and service terms: “What does X’s Y service include, and how much does it cost?” Focus on where the numbers it cites were copied from (we got burned once: an example price from an explainer article three months earlier was treated as an official quote).
- Layer three: category terms: “Which X service providers are there in Beijing? Recommend a few.” This layer matches how real new customers actually ask, and it is also the hardest.
For each layer, fix three platforms, fix the wording, and fix one day each month. Screenshot the answers and archive them, and record the cited sources as well.
What you record is a trend, not a score. A single “one hit out of three platforms” proves nothing. But if three months later it becomes two hits, or the cited sources shift from directory listings to your own website and established media, then it really is moving in the right direction. Conversely, if nothing has moved for six months, the current approach is pointed the wrong way and it is time to change tack.
Four things not to do
- Don’t test brand terms only. It is the easiest layer to pass and the one that affects customer acquisition least, so passing it easily breeds false confidence.
- Don’t test on just one platform. The table above is the answer: on the same day, the conclusions can be completely opposite.
- Don’t look only at the answers and ignore the sources. The sources tell you what to do next; the answers only tell you whether you are OK right now.
- Don’t treat a single change as cause and effect. AI answers are random; changing one word in the question can change the result. Look at the trend, not at one instance.
A few common questions
Q: Does testing this cost money?
A: No. You can just ask directly on the web version of mainstream AI platforms, and it takes about ten minutes each time. What it costs is the discipline of keeping up the monthly records, which is harder than the technical barrier.
Q: How long does GEO take to show results?
A: Months. AI has to re-crawl and re-trust the sources, so it cannot be fast. But precisely because it is slow, you need baseline records all the more; otherwise you cannot tell whether it is “not working yet” or “the direction is wrong.”
Q: If the brand terms come back correct, does that mean GEO is done well?
A: No, that is the most common misjudgment. Accurate brand-term answers usually only mean your website is clearly written and your business registration information is public. What actually brings in business is the category-term layer, and that layer is an order of magnitude harder.
Q: Are small companies completely out of luck on category terms?
A: Large players do have a natural advantage on generic terms, with analyst reports and market share behind them, so taking them head-on is pointless. The direction should be toward specific scenarios rather than fighting for the broadest term. Still, I have to honestly add one counterexample here: for the same scenario-specific question (“A 50-person company wants to outsource IT entirely and not buy equipment itself. Who should it call in Beijing?”), we got a hit five days ago, but no hit this time. That actually proves the point above: the noise in a single result is larger than the signal, so don’t treat one test as a conclusion. Look at the trend over several months.
Want to know how AI describes your company right now, and where you should invest?
Siwenlide runs website and SEO/GEO content-driven customer acquisition for small and mid-sized businesses in Beijing and the surrounding area: baseline measurement across multiple AI platforms using the three-layer question set, with the cited sources recorded alongside, to judge whether you need to strengthen your website, your media endorsements, or your comparison content, then a monthly comparison report so you can see the trend. We run this measurement on our own site first, and the data in this article is the real result of one such round. Visit siwenlide.com (or start with our company profile), or call 400-686-2011 to talk it over.
