Last month I was doing a traffic review with a client who runs a foreign trade tool site, and he asked me a very practical question: for the same keyword, we rank #2 on Google, but clicks in the last two months are down 30% compared to the same period last year. Rankings haven’t dropped, but clicks have. After digging around, the answer was hiding in that AI overview at the top of the results page — the user’s question got answered directly by AI, our link sits below, and nobody scrolls down to see it.
This isn’t an isolated case. Starting in 2025, Google’s AI Overviews gradually covered more queries, and ChatGPT Search and Perplexity became many people’s default search entry points. The underlying logic of keyword research has genuinely been rewritten. This article talks about the practical changes I’ve made over the past few years: which old methods still work, which ones need a new approach, and what my current keyword selection process looks like.
First, admit one thing: search volume numbers are getting less and less reliable
The traditional keyword selection logic is simple: find high-volume keywords, assess difficulty, write content, wait for rankings. This logic works on one premise — that search volume converts into click volume. But after AI overviews appeared, this premise started to wobble.
Take a definitional query like “what is keyword clustering” as an example. When a user types this, what they want is a one-sentence definition. Before, that definition required clicking into a page to see; now AI answers it directly at the top of the results page, and顺便 lists our site name as a citation source. The user got the answer but didn’t click our link. Impressions still count, but clicks go to zero.
I’ve crunched the data on several sites I manage: for definitional, comparison, and simple step-by-step queries, click-through rates generally dropped 20%-40% after AI overviews appeared. This isn’t because SEO was done wrong — it’s because the traffic distribution rules changed. Still using the old logic for these types of keywords is like pouring water into a leaky bucket.
To judge whether a keyword is still worth doing, I only look at three things
Since search volume can’t be fully trusted, I now run keywords through three checks first, and the order can’t be reversed:
- First, when you search this keyword, is there an AI overview in the first screen of results? If so, does AI answer it completely? If AI finishes the question in one sentence, clicks for this keyword are basically locked down — unless you can get cited by AI.
- Second, is the search intent “wanting an answer” or “wanting a method”? The former easily gets packaged into an overview; the latter requires users to click in for the full walkthrough, so clicks are safe.
- Third, does the answer for this topic have “citation potential” — i.e., is your content clearly structured, with explicit conclusions, that AI can excerpt into a paragraph? This determines whether you go from “getting traffic cut off” to “getting cited.”
Only after passing all three checks is a keyword worth writing about. If one doesn’t satisfy, first think about whether the content format can be adjusted — if not, set it aside.
How to quantify how much a keyword gets cut off by AI
Open GSC’s “Queries” report, pull out keywords with high impressions but CTR below 2% — these mostly appear on results pages covered by AI overviews. Then multiply the tool’s search volume by your real CTR to get the actual clicks you can get. For example, a keyword the tool shows as 12,000 monthly searches with a real CTR of only 1.5% lands you about 180 visits — that end number is closer to the truth than search volume. To verify further, search 3 core keywords in incognito every week, screenshot whether AI overviews give the full answer, and after four consecutive weeks you can chart the “traffic cut-off ratio” trend.
How I do keyword selection now: a three-step process
I’ve been running this process for almost a year, and the core is one sentence: shift from “selecting keywords” to “selecting questions.” The steps are simple, but each one has its nuances.
Step one, break down intent. Classify candidate keywords by “wanting an answer” vs. “wanting a solution.” For answer-seekers, create structured content to compete for citations; for solution-seekers, create deep tutorials to compete for clicks. I’ve seen too many sites mix these two together, ending up with something that’s neither here nor there and misses both ends.
Step two, check citation potential. Open the results page, see who AI overviews currently cite and which passage they cite. If it’s citing the #5 page, that position has opportunity; if AI doesn’t even show an overview for this query, the keyword is still safe for now.顺便 open the cited page and study how its opening is written — those 100 words are the “standard answer template” in AI’s eyes.
Step three, calculate real value. Stop staring at the tool’s monthly search volume — go to GSC and look up this keyword’s real clicks over the past three months. Then search the keyword once more and estimate how much AI overviews are eating. Real clicks minus AI cut-off is the increment you can争取. After doing this calculation, many “high-traffic keywords” get crossed off my list directly — the ROI just doesn’t add up.
In the AI era, what kind of content gets cited
After six months of AI citation experiments, I’ve总结出 three common traits of cited content, for your reference:
- Give a clear conclusion within the first 200 words. AI likes “conclusion first, then展开” structure because that’s what it excerpts.
- Opinions backed by data or cases. Cited content is usually “verifiable” — with specific numbers, real cases, sources. Vague “best practices” nobody cites.
- Clear paragraph structure, one point per sentence. AI excerpts by semantic chunks — the clearer the structure, the higher the chance of getting cut out.
When rewriting, put each article’s core conclusion in the first 150 words, use H3 to break down steps, mark sources for key data. Don’t stuff keywords — AI cares more about whether you answered the question. Write users’ likely follow-up questions as an FAQ section, covering “what is it, why, how to do it”; laying out synonyms in advance lets you hit more phrasings.
Honestly, direct clicks from AI citations are few, but brand exposure is real. A user asks AI “how to do keyword clustering,” AI’s answer mentions your site name, and the user remembers you. Next time they search for a specific question and you appear in the results, the click probability is completely different. This is a long-term play — you have to count it.
Citation potential checklist
Treat the table below as a pre-publish checklist — only publish after passing all four. Review once a week, treat “conclusion first” and “data sources” as hard thresholds. Do this consistently, and the rate of getting hit by AI overviews will steadily rise.
| Check item | Passing standard | Common deduction points |
|---|---|---|
| Conclusion first | Give answer in first 150 words | Three screens of buildup before getting to the point |
| Data sources | Key numbers marked with sources | No citations throughout |
| Clear structure | H3 breaks down steps | One big block of text |
| Semantic coverage | Includes synonym variants | Only stuffing core keyword |
Three suggestions for people still using old keyword methods
If you’re still selecting keywords sorted by search volume, I suggest doing three things starting today:
- First, pull real clicks from GSC and compare with tool search volume, find keywords with “high impressions, low clicks” — these are the hardest-hit by AI cut-off, mark them first.
- Second, pick 2-3 core keywords every week, search in incognito to see results page format, screenshot and archive. Compare after a month, and changes in AI overview coverage become obvious at a glance.
- Third, for newly written articles, write the conclusion in the first 200 words to make content “citable.” This isn’t about pleasing AI — it’s about letting users know in 10 seconds whether this article is useful to them.
Tools won’t disappear, but the people doing keyword research need to change their thinking. Search volume is the past, citation potential is the future. I’ve been saying this for a year, and looking back, the direction is right.
