Ask AI for topic ideas and it’ll give you a hundred in a minute. I tried it once, and my takeaway was: half are common knowledge, a quarter are nonsense, and of the remaining quarter, barely any are usable. AI is good at breadth, bad at judgment — it doesn’t know which keywords people actually search for, or which topics you have the ability to win at.
So the key is building a validation process to run each AI-generated topic through a filter. This article breaks down the “four validation questions.” Use this standard to screen AI topics, and you’ll block out 80% of the junk — what’s left is genuinely worth writing. After our team started using this, our topic pass rate stabilized at around 30%, up from under 10% when we were writing blind — the writing manpower we saved effectively doubled.
Figure: Four-question funnel for validating AI topics (compiled by 运营GO)
The four validation questions
1. Is anyone actually searching for this? Open GSC and check related queries, or use a keyword tool to see search volume. A lot of AI-generated topics have zero search volume — writing them is just self-indulgence. A quick trick: throw the AI topic into GSC’s “Queries” report and search — if there are impressions, it means Google already recognizes your site as relevant to that keyword, and the ranking cost of writing it will be lower.
Don’t get excited just because a tool says “1,000+ monthly searches.” The tool measures the overall market — whether your site can capture a slice of it depends on your current authority and the competitive landscape at the top. I trust real impressions in GSC more — that’s hard evidence that my site is already recognized as relevant, stronger than any external estimate.
2. What problem is the searcher trying to solve? Judge the search intent. For example, “how long does SEO take to work” is an anxiety-type question — users want a timeline expectation and real cases; “SEO tool comparison” is a decision-type question — users want an objective comparison. The content format follows the intent — what structure to use, whether to include tables, all determined by this question. Get the intent wrong, and no matter how well the article is written, it won’t catch the reader. For a systematic explanation of intent judgment, see the keyword intent mapping funnel article, which breaks down the writing differences between informational, commercial, and transactional types in detail.
Intent also determines “how deep to write.” Informational questions just need a clear concept and steps; commercial investigation needs comparison tables, pros and cons, use cases; transactional directly gives a purchase decision checklist. AI-generated topics often don’t label intent — this step must be done by a human, because if the intent is wrong, the whole content is wrong.
3. Can we outrank the competition? Look at the top 10 search results. If they’re all industry giants and big UGC platforms, a new site going head-to-head basically has no chance. If there are small-to-medium sites ranking in the top results, it means there’s an opening — that’s also why competitor keyword filtering goes back to GSC to check if small sites are in the top results (the third layer of that funnel is exactly for this).
4. Can we convert the traffic after writing? When traffic comes, can we convert it? Do we have a corresponding product, landing page, or email sequence to catch it. A lot of topics have decent search volume but are far from money — writing them only inflates vanity metrics. Prioritize topics where “the searcher’s next step can use what we offer.” A judgment trick: imagine this article ranks #1 — what will the visitor do after reading? If the answer is vague, this topic isn’t worth writing right now.
Four-question quick reference
| Question | Fail signal | Pass signal |
|---|---|---|
| Anyone searching? | Tool shows zero volume, no GSC impressions | GSC has impressions or tool shows 300+ monthly searches |
| Intent clear? | AI topic has no intent label, feels like common knowledge | Can articulate what content format the user wants |
| Can we outrank? | Top results are all giant sites | 1-3 small-to-medium sites in top results |
| Can we convert? | No corresponding product/landing page | Has product page or email sequence to catch traffic |
How to turn the four questions into a process
Don’t run through this in your head — write it into a spreadsheet, that’s most reliable. Each row is one AI topic, four columns for the four questions, fill in “pass/fail” plus a one-line reason, mark fails in red. Spend an hour each week screening a batch — the pass rate is usually only 20-30%, and the rest get tossed directly. This process itself is valuable — it helps you separate “looks busy” from “actually worth writing.”
There’s also an advanced use: ask AI not just for topics, but also for its guess on “why people would search for this topic,” then you take that to GSC to verify. AI’s guesses can often point you in a direction, but the conclusion must be decided by you with real data — don’t treat AI’s “people probably search for this” as fact. AI is the expander, humans are the filter — this division of labor can’t be reversed. Letting AI judge whether you can outrank is the biggest pitfall.
Boundaries of AI use
Let’s be clear about what AI can and can’t do, so the team doesn’t misuse it. AI is good for: brainstorming to find blind spots, expanding one topic into multiple angles, supplementing synonyms and long-tail variants. AI is not good for: judging search volume (it has no real-time data), judging whether you can rank (it doesn’t know your authority), judging conversion value (it doesn’t understand your business). These three things must be done by humans, or with GSC and tools. Draw the boundaries clearly, and AI can be a helper instead of a liability.
Here’s a real screening example. AI once gave 20 “content operations” topics — after the four questions, only 5 made it into the pool: “content calendar template,” “how to write short video scripts,” “how to cold-start B2B content,” “how to package user case studies,” “what to do when WeChat open rates drop.” Of the 15 tossed, 8 had zero search volume (“future trends in content operations”), 4 we couldn’t outrank (“viral copywriting formulas” — top results all giants), 3 we couldn’t convert (“which content operations certificate to get” — unrelated to our business). Each of the 5 kept topics maps to an existing product capability we have — write it and it can catch conversions.
This example illustrates one thing: AI’s value isn’t “giving answers,” it’s “giving you a wider net” so you miss fewer angles. But the final call is always data and business judgment. Treat AI as the expander and humans as the filter, and topic quality will be higher than either pure manual or pure AI alone.
