--- title: "Keyword Research In The Age Of AI Search | Gudu" description: "Search volume told you what people typed, not what traffic is left. How to research questions and entities now that AI answers resolve the intent." canonical: https://gudu.com.au/keyword-research-in-the-age-of-ai-search/ --- # Keyword research doesn’t work the way it used to By Bianca Hoare · September 10, 2026 The keyword tool still returns a spreadsheet. What’s changed is what the spreadsheet is worth. A list of phrases with monthly volumes beside them used to be a plan. You picked the winnable ones, wrote a page for each and waited. That worked because a search meant a results page, and a results page meant somebody had to click something. Both halves of that sentence have come loose. ## Where the click went Ask Google a question now and there’s a decent chance it writes you an answer before the ten links begin. Ask an assistant and there are no links at all unless it decides to cite some. The search still happens. The visit often doesn’t. So a volume figure has quietly stopped meaning available traffic. It means interest. Some of that interest still ends in a visit, and it’s worth knowing which kind: people comparing options, checking a price, working out whether to trust a business, or getting close enough to buy that they want to see who they’re buying from. Those searches still come looking for a website. Definitional ones increasingly don’t, which doesn’t make them worthless, it makes them a different job. You write those to be quoted rather than to be clicked. ## Volume was always a proxy It measures how often a phrase was typed into one search engine. That’s all it ever measured. It never told you what a customer was worth, whether the searcher was a buyer or a student or a competitor doing exactly what you’re doing, or whether the phrase resembled how anyone actually speaks. Now it also misses every question put to an assistant, because none of them publish query counts. Anyone quoting you a monthly volume for ChatGPT made it up. The other thing volume hides is how contested a term is. In Australia, Ahrefs puts “ai seo” at roughly 1,600 searches a month with a difficulty score of 68, and “ai seo agency” at about 900 with a difficulty of 1 (Ahrefs, August 2026). Same subject, and the smaller number is the one you can win. Anyone sorting a keyword export by volume and starting at the top will pick the wrong row nearly every time. ## Research the question, not the phrase Before I open a tool now, I build a question list from places that have nothing to do with search. The sales team’s repeated explanations. The inbox. The site’s own search box, which is the most underread report in most businesses. Search Console’s query list, which shows what you already show up for rather than what you hoped. The people-also-ask box. Whatever forum or subreddit your customers actually use. Then I write each one out as the sentence a person would say, not as a keyword string. What comes back tends to be the same handful of concerns wearing different clothes. “How much does a website cost in Australia” and “why are two website quotes so far apart” are one customer at two moments of the same afternoon. They belong on one page, and that page will get found for both. ## Cluster by meaning, not by string The old habit was a page per phrase, which produced near-identical pages competing with each other for the same result. Engines resolve synonyms perfectly well and have for years, and the systems writing answers do it even more aggressively, because they’re matching meaning rather than looking anything up in an index of strings. So write one page that answers the question properly and let the variants live inside it. Keep Australian spelling, since that’s what your customers type and it signals where you operate. If a term genuinely splits into two different intents, split the page. “SEO packages” and “what SEO includes” are the same intent. “SEO” and “SEO for Shopify” are not. ## What a brief looks like now One page, one question, and the sub-questions that come with it. Ask the question in the subheading and answer it in the next forty to sixty words in sentences that stand up alone, without relying on the paragraph above. Elaborate underneath for as long as it deserves. Use a real table when you’re comparing things, with real header cells, because a machine reads a table far more reliably than four paragraphs of comparison prose. Then put in the part nobody else can copy. Your numbers, your pricing logic, what you see across your clients, dated and with the method attached. That’s the material that gets cited, and being cited is now part of what a page is for. We’ve written the fuller version of that argument in [what generative engine optimisation actually is](https://gudu.com.au/what-is-generative-engine-optimisation/). ## Where volume still earns its keep It settles arguments about framing. If your team says “digital strategy” and the market says “marketing strategy”, the data tells you who to write for. It also sizes decisions. A topic with real volume might deserve its own page, and one with almost none is usually a section inside another. And on a smaller site, difficulty matters more than volume does, because a term you can’t reach is worth nothing regardless of how many people search it. Australian search results are full of soft doors: “seo packages” has about 1,200 searches a month at a difficulty of zero (Ahrefs, August 2026), which is why we bothered writing [what SEO costs in Australia](https://gudu.com.au/what-seo-costs-in-australia/) rather than chasing the head term above it. ## Measuring what you can’t see Search Console remains the least bad instrument you have. Appearances inside AI Overviews mostly sit within the impression counts, so a page whose impressions climb while clicks stay flat is often being read inside an answer rather than ignored. That pattern used to look like failure. Now it needs checking before you conclude anything. AGL is the clearest version of this we’ve worked on. Over six months our SEO team saw a 58% increase in average search position, 40% more keywords sitting in the top three, and 14.5k features in Google’s AI Overviews, on a site where a large share of the audience never needs to leave the results page. The [AGL case study](https://gudu.com.au/agl/) has the detail. Watch branded search volume and direct traffic alongside it, since being named inside an answer often shows up later as somebody typing your name. And ask new enquiries what they searched. Self-reported data is messy and it’s occasionally the only evidence you’ll get. ## What I’d change first Export your top twenty keyword rows and rewrite each one as a question. Half will collapse into each other. Merge the pages that were competing for the same answer, and give the survivors a proper subheading and a direct answer underneath. That’s a morning’s work and it’s the highest-value thing on this list. Everything else is a programme. If you want help running it, that work sits across our [SEO](https://gudu.com.au/services/technology/seo/) and [GEO](https://gudu.com.au/services/technology/geo/) teams, and the practical habits behind it are in our guide to [improving your SEO](https://gudu.com.au/how-to-improve-your-seo/).