How Content Websites Can Actually Survive AI Overviews
AI Overviews and AI Mode are cutting clicks to informational pages. Here's the substitution-risk framework I use to decide what to fix, cut, or rebuild.
Quick answer
AI Overviews and AI Mode really do reduce clicks on informational search results, and the research behind that is solid enough that arguing about it is a waste of time. But the effect isn’t evenly distributed. It hits hardest on pages whose entire value fits in a paragraph: definitions, generic tutorials, “best of” roundups built from other people’s research, basic conversions, thin FAQ content. It hits much less on pages that need something an AI answer can’t reproduce: original data, first-hand testing, an interactive tool, live information, or a specific source someone already trusts.
The useful response isn’t chasing “AI SEO” tricks. It’s auditing your content by how easily each page can be substituted, cutting investment in the pages that can’t win that fight, and putting the saved effort into things that either resist summarization or convert a search visitor into someone who returns without needing Google again. That’s the rest of this post, with the research behind it.
The bargain that’s breaking
For most of the web’s commercial history, the deal was simple. You wrote something useful, Google indexed it, and if you ranked, Google sent you a person who read your page, saw your ads, clicked your affiliate link, or joined your list. You didn’t own the discovery channel, but you got paid for renting it in the currency that mattered: visits.
AI Overviews and AI Mode change those terms without renegotiating them with you directly. When Google, Bing’s Copilot, ChatGPT, or Perplexity can read your page, synthesize it with a dozen others, and hand the user a complete answer inside the results, the visit that used to be payment for your work often just doesn’t happen. Google still crawls your site and may still cite you. It increasingly doesn’t need to send anyone to you to deliver the value your page created.
Cloudflare, which sits in front of a large share of the web’s traffic, tracks what its crawlers see. In an August 2025 analysis, it found Anthropic’s crawlers fetching roughly 38,000 pages for every one referral visit, OpenAI around 1,091:1, Perplexity near 194:1. A year later, the pattern hadn’t reversed, it had shifted composition: by June 2026, 52% of AI crawler requests were for model training, up from 22% in spring 2025, while the share used for live search kept shrinking. Cloudflare’s response has been to start blocking AI training crawlers by default on new domains unless the site owner opts back in, an infrastructure-level bet that a healthier market needs scarcity and consent, not just more crawling.
Google disputes the framing that this broadly hurts publishers. In an August 2025 post, Search VP Liz Reid wrote that “total organic click volume from Google Search to websites has been relatively stable year-over-year,” with no supporting data published alongside the claim. Independent measurement tells a sharper story, and specifically contradicts the idea that any stability is evenly shared. Chartbeat data reported by Axios in March 2026, covering roughly two years across thousands of client sites, found referral traffic from search down 60% for small publishers (1,000 to 10,000 daily page views), 47% for medium publishers, and 22% for large ones, with Google Search page views specifically down 34% between December 2024 and December 2025. Chatbot referrals grew over 200% in the same window and still made up under 1% of total publisher page views. Aggregate stability, if real, is entirely compatible with that shape: clicks consolidating toward a smaller set of large domains while smaller, independent sites absorb most of the loss.
I don’t know your specific numbers, and if you’re running a smaller independent site, the data above suggests you should assume you’re on the more exposed side of that gap until your own numbers say otherwise. Either way, the useful question isn’t whether this is happening industry-wide. It’s which of your pages are exposed, and what to do about them.
What the click data actually shows
Two studies get cited constantly, usually stripped of methodology.
Pew Research Center tracked the actual browsing behavior of roughly 900 US adults for March 2025 via a consented browser tracker: 68,879 unique Google searches, 12,593 of which produced an AI summary. When a summary appeared, users clicked a traditional result in 8% of visits, versus 15% without one. Clicking a link inside the summary itself happened in just 1% of visits. Searches with a summary were also more likely to end the session entirely (26% versus 16%). Pew is explicit about the limitation: this covers Google only, and can’t fully separate the summary’s effect from query type, since simple factual queries are both more likely to trigger a summary and more likely to have been satisfied by one anyway.
Ahrefs took a different approach: aggregated Search Console data across 300,000 keywords, isolating the AI Overview effect by forecasting what click-through rate the AI-Overview keyword group would have reached without the feature and comparing that to what actually happened. Its first run, comparing March 2024 to March 2025, put the effect at roughly a 34.5% reduction in position-one click-through rate. Ahrefs reran the same methodology in February 2026 on fresher data, December 2023 versus December 2025, and the number had nearly doubled: an average 58% reduction in position-one CTR (down from 7.6% to 1.6%), with the drop fading by position (position two down 50.8%, position ten down 19.4%). The direction matters more than either single number: the effect isn’t a one-time adjustment that plateaus, it’s been getting worse as the feature matures and users adapt to it. A separate Ahrefs study of what triggers an AI Overview, covering 146 million SERPs from September 2025, found overviews on 20.5% of all queries, but that varies hugely by shape: 9.5% of single-word queries versus 46% of queries with seven or more words, 57.9% of question-phrased queries versus 15.5% of non-questions, and 99.9% informational intent. Topic matters too: science (43.6%) and health (43.0%) queries trigger overviews constantly; shopping (3.2%) and real estate (5.8%) almost never do.
Together these point at a pattern worth internalizing: AI Overviews concentrate on exactly the query shape commodity informational content lives on (long, explicit, informational, no clear brand or transaction attached), and when they appear, they measurably absorb the click an organic result would have gotten. That’s two independently run studies, different methodologies, landing on the same shape of effect. What neither proves is that this is uniform or permanent for your business specifically. Query mix varies enormously by niche, and “fewer clicks on average” is a population statistic, not a verdict on your page.
Three things people lump together
A lot of “AI SEO” advice, including plenty written on exactly this topic, confuses three separate questions:
- Can search engines find, understand, and rank your page? Traditional search visibility. Hasn’t fundamentally changed.
- Does an AI system use, mention, or cite your page in an answer? AI visibility. Genuinely new.
- Does any of that produce a visitor, subscriber, customer, or revenue? Publisher economics. The only one that pays the hosting bill.
Content strategy that maximizes #2 while assuming it automatically produces #3 is common right now, and wrong. Being the quiet source an AI answer draws from, with no click and no attribution the user notices, can maximize AI visibility while doing almost nothing for your business. Citation is not a business model. More on that below.
What Google actually says to build
Google published its first consolidated guide to optimizing for generative AI features in May 2026. The headline claim is that there’s no separate ranking system to game: “the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems.” The part that matters more is the distinction Google draws between commodity and non-commodity content.
“Commodity content (for example, something like ‘7 Tips for First-Time Homebuyers’) is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers.”
Non-commodity content, by contrast, reflects “unique expert or experienced takes that go beyond common knowledge and the ordinary,” where “a first-hand review provides a unique perspective based on personal experience, whereas a summary of existing content simply restates information already available elsewhere.” Read plainly: Google is describing, in its own words, exactly the content least likely to matter in an AI-mediated result, and it happens to line up almost exactly with what’s easiest for an AI system to compress into a paragraph.
The same document is unusually blunt about tactics sold constantly as “AI SEO.” On llms.txt files: you don’t need them to appear in Search, and they “neither harm nor help your site’s visibility or rankings.” Google added a specific clarification in June 2026 after Chrome’s Lighthouse started surfacing an “agentic browsing” audit that checks for the file’s existence, which caused a wave of people to assume its presence had suddenly become a ranking signal. It hadn’t; Search still ignores the file. On content chunking: “there’s no requirement to break your content into tiny pieces for AI to better understand it.” On rewriting for AI: “AI systems can understand synonyms and general meanings.” On structured data: “there’s no special schema.org markup you need to add” for generative AI features specifically.
None of that makes schema, clean HTML, or fast pages pointless, they’re the same technical hygiene they always were. It means they aren’t a new lever specific to AI Overviews. Time spent on an llms.txt file instead of the content itself is optimizing for a control Google has told you, in writing, does nothing.
The content AI actually threatens
Put Google’s commodity framing together with the Pew and Ahrefs data, and a clean line appears. AI answer engines disproportionately threaten content whose entire job is: I collected information that already exists elsewhere, and reorganized it.
Higher substitution risk: plain definitions and “what is X” explainers, basic factual questions with one settled answer, generic tutorials indistinguishable from fifty others, unit conversions and quick reference facts, list posts synthesized from the same top-ten results everyone scraped, thin affiliate roundups built from spec sheets, templated long-tail pages built to catch a keyword rather than answer a real question.
None of that means these categories are dead. It means their substitution risk is structurally higher, because the entire value fits in the paragraph an AI system is already generating. Contrast that with content that has to be experienced rather than summarized: original experiments and first-hand testing, proprietary datasets or surveys, interactive calculators and generators, searchable databases and comparison engines, frequently changing or live information, community discussion, original photography and video, deeply opinionated and accountable expert analysis, hyper-local or highly specialized knowledge.
The distinction isn’t long versus short, or expert versus amateur. It’s whether a page’s value can be read or whether it has to be experienced, queried, verified, downloaded, watched, or trusted. AI systems are very good at the first kind of value and structurally incapable of the second, because producing it requires actually running the experiment or holding the dataset, not describing what one would look like.
The substitution test
This isn’t a ranking-factor checklist, and it won’t predict your position in Google. It’s a business-risk audit: a way to estimate how exposed a specific page is to being replaced by a paragraph the user never clicks through to reach. Pull up one real page on your site, ideally one that already earns you traffic or revenue, and score it honestly against the eight signals below.
Score each signal 1 (low risk) to 3 (high risk) for one real page on your site:
Can the core answer fit in a single paragraph?
Is this exact information available on dozens of other sites?
Is there little to no original data or first-party research behind this page?
Could this page's writer have skipped any first-hand experience and produced the same result?
Does the visitor get the full value without interacting with anything, no tool, search, or input?
Is there little reason for someone to come back to this specific page again?
Would most readers accept any competent source instead of specifically yours?
Would quoting three sentences from this page hand away most of its value?
Your score
–/ 24
Score each item above to see your result.
- 8–13Your result
Low substitution risk. Original data, interaction, or a specific enough source that a paragraph can't replace it. Defend this page and keep it current.
- 14–19Your result
Medium substitution risk. Competent and useful, but not distinct enough yet. Add a first-party angle before assuming it's safe.
- 20–24Your result
High substitution risk. This page's whole value could fit in an AI answer box. Treat it as a candidate to upgrade, consolidate, or stop investing in.
The table below is the same eight signals as a quick reference, useful for scanning a longer list of pages without scoring each one individually:
| Signal | Higher risk | Lower risk |
|---|---|---|
| Can the answer fit in a paragraph? | Yes | No |
| Is the information available on many other sites? | Yes | No |
| Original data or first-party research | None | Significant |
| First-hand experience required to write it | Not really | Essential |
| User interaction needed to get the value | No | Yes |
| Recurring reason to revisit | No | Yes |
| Brand-specific demand (“I want your answer”) | Low | High |
| Three quoted sentences would give it away | Yes | No |
Neither tool predicts rankings, and a low-risk page can still rank poorly for unrelated reasons. It’s a lens for where to spend limited time, not a replacement for SEO fundamentals.
Stop publishing information. Start originating it.
The highest-leverage shift for an existing content site is moving budget away from pages that only collect information and toward pages that create it. A generic “Best Hiking Shoes” roundup is commodity content: specs pulled from manufacturer pages, arranged into a list, competing against a hundred sites that did the same research the same way. Compare it to “I Walked 500km in Six Hiking Shoes, Here’s What Actually Wore Out”: the underlying data (wear patterns, specific failure points, mileage before a sole delaminated) doesn’t exist anywhere else until you produce it. An AI system can summarize your conclusion. It cannot retroactively walk the 500km for you.
I did something adjacent with the Nuxt vs Astro benchmark comparison on this site: rather than write another “which framework is better” opinion piece synthesized from other people’s takes, I built the same app in both frameworks and measured the actual bundle-size difference. An AI system can summarize that conclusion accurately. It can’t run the build itself, and it can’t tell me what changed after the next major version, because that requires doing the work again.
Be honest about one thing, though: originating information doesn’t automatically translate into traffic, even when it gets cited. If your research gets absorbed into an AI answer alongside everyone else’s sources, you may get a citation and zero incremental visits. That’s not a reason to skip the original work. It’s a reason not to treat getting cited as the finish line.
Citation isn’t a business model
There’s a real chain of value here, and it gets more economically meaningful the further right you move: citation → visibility → visit → relationship → conversion → revenue.
Being cited by an AI system, assuming you can even verify it, sits at the weak end of that chain. Pew’s data is direct on this: even when a summary cites sources, users click into them in just 1% of visits. Google’s generative AI performance reports in Search Console, rolling out from June 2026, illustrate the same gap structurally: the report shows impressions inside AI Overviews and AI Mode by URL, country, and device, but at launch excludes clicks, click-through rate, the actual queries, and average position, and it’s only reaching a subset of sites. You can sometimes see that you were shown. You still can’t reliably see whether anyone acted on it.
Google has also been rewarding sources users actively choose. Preferred Sources, which lets users hand-pick publishers to prioritize, expanded from Top Stories into AI Overviews and AI Mode in May 2026, with links from a user’s preferred sites reportedly getting roughly double the click rate of unlabeled links. Notice what that actually rewards: being a source someone already trusts enough to select, not being cited passively by an algorithm. It reinforces the case for brand, not a substitute for it.
This isn’t a Google-only gap. Microsoft’s Bing Webmaster Tools shipped its own AI Performance report in February 2026, tracking how often a site is cited as a source across Copilot, AI-generated summaries in Bing Search, and select partner surfaces, then expanded it in June 2026 with intent and topic groupings and a “citation share” metric showing your slice of citations against competing sites for a given query. It’s a genuinely more detailed citation dashboard than Google’s equivalent at this point, and it still reports citations only, no clicks, no CTR, same gap as Search Console’s report. One publisher’s account of the practical result of that gap, described in independent reporting on Bing’s AI Performance data, put it starkly: content cited roughly 48,000 times across Copilot and Bing’s AI summaries over several months, against roughly a dozen resulting clicks back to the site. That’s one site’s number, not a benchmark to extrapolate from, but it’s the same shape as Pew’s 1% figure for Google: citation volume and referral traffic aren’t just loosely correlated, they can be almost entirely decoupled.
ChatGPT shows a variant of the same gap. Similarweb’s 2026 Generative AI Landscape report found that 65% of the URLs ChatGPT cites sit two or three folders deep, exactly the specific, substantive pages this article keeps arguing are worth building. But 58.8% of the actual referral traffic ChatGPT sends to the open web lands on homepages, not on those cited pages. Being the deep, specific source an AI system draws from and being the page a human actually opens turn out to be two different outcomes, and right now the second one mostly isn’t going to the pages doing the work.
Visibility tracking isn’t worthless as a leading indicator. Just don’t let “we’re cited a lot in AI Overviews” become the metric your business reports on. A strategy that maximizes citations while producing almost nothing on the visit-relationship-revenue side of that chain is closer to a vanity metric than a growth channel.
Build things that can’t be summarized
If commodity information is the most exposed asset a content site owns, the natural response is to own less of it and more of something an AI system can’t reproduce by reading your page: a tool. A mortgage calculator can’t be summarized, because its value only exists once someone enters their own numbers. Neither can a compatibility checker, a searchable product database, a price tracker, a validator, a generator, or a checklist that remembers where you left off. These aren’t content in the traditional sense; they’re small pieces of software wrapped in a page, and their value depends on interaction, which a text summary can’t provide.
This is a real structural advantage for developers who also publish, over publishers who only write. Building a working calculator or a small searchable dataset is a software decision, not a content one, and it’s straightforward for someone who can already script or wire up an API, while representing a genuine barrier for a team that only produces prose.
I don’t want to oversell this. Tools aren’t a permanent moat. Agentic browsers are already being built to fill in forms and interact with pages on a user’s behalf; Google’s own guide to building agent-friendly websites, published in early 2026, treats agents as a distinct visitor type that inspects your DOM and accessibility tree. A sufficiently capable agent could eventually learn to operate a simple calculator and report the result back without a human seeing your page. Tools aren’t immune to automation forever. They move you further from the specific failure mode of commodity text (a paragraph absorbed wholesale into someone else’s answer) and closer to something that needs a live interaction with your actual product, which is a meaningfully harder thing to substitute away.
Turn a search visitor into someone you don’t have to find again
Most content sites never converted an anonymous search visitor into anything. They got the click, showed the ad, and the relationship ended there, meaning every visit depended on ranking again, on that same query, forever. That was always a fragile model; AI search just makes the fragility visible faster.
The fix isn’t “start a newsletter” as a generic instruction, because a newsletter with no reason to exist gets no signups. “Subscribe for our latest articles” is a request for attention with nothing offered back. What actually converts is built around a specific, recurring trigger: an alert whenever a specific reference rate or regulation changes, rather than a general finance newsletter; a weekly digest of flight deals from one airport, rather than “travel tips”; a price-drop notification for one category someone is actually watching; a saved search that emails you when new results match, rather than asking you to search again.
The mechanism matters less than the discipline: converting one Google click into a relationship that no longer depends on Google every time. Accounts, watchlists, RSS feeds, community features, even an active presence on YouTube or a niche forum do the same underlying job. They give someone a reason to come back directly instead of rolling the dice on whether an AI system answers their next question before you get the click.
What query fan-out actually changes
One substantive shift behind AI Mode, easy to either overhype or dismiss, is what Google calls query fan-out. When someone asks a complex question in AI Mode, the system breaks it into multiple related sub-queries, runs them in parallel against Google’s index, and synthesizes the results, potentially pulling in sources for aspects of the question the user never typed. Google’s own AI features documentation confirms the mechanism, describing AI Overviews and AI Mode as surfacing “a wider and more diverse set of helpful links” than a standard results page.
The practical shift isn’t “write more pages to catch more sub-queries.” Traditional keyword SEO asks: what’s the exact phrase people type, can I rank for it? Fan-out-aware thinking asks: if someone asked a bigger, messier version of this question, what specific piece am I the strongest possible source for? A query like “which resort in Andorra works best for a beginner traveling with kids in February without renting a car” decomposes into resort difficulty, childcare availability, snow reliability, transport from the airport, and lift pricing. A site with a genuinely detailed, first-hand page on any one of those sub-questions has a shot at contributing to that answer, without ever ranking for the compound query itself.
Be careful with the conclusion, though. The temptation is to mass-produce thin pages targeting every conceivable fan-out variant, and Google’s spam policies explicitly define exactly that (generating many pages primarily to manipulate rankings rather than help users) as scaled content abuse, regardless of whether AI, humans, or a mix produced the pages; Google formally extended these policies to cover the AI-generated layer of Search itself in 2026. The useful takeaway isn’t “produce more pages.” It’s “go deeper on the specific sub-questions you’re genuinely positioned to answer better than anyone,” a smaller and more defensible target than chasing every keyword permutation.
Sorting an existing content portfolio
Treating a few hundred existing posts as individual crises is a good way to burn a month doing nothing useful. Sort the portfolio into buckets and decide per bucket:
| Category | What it looks like | Action |
|---|---|---|
| Strong original asset, real traffic and revenue | Unique data, tool, or expertise; converts well | Defend, keep it current |
| Commodity page, still gets traffic | Ranks fine today, easily summarized | Upgrade with first-party data or interaction |
| Commodity page, little traffic | Low value either way | Consolidate or retire |
| High impressions, collapsing click-through rate | Still shown in search, fewer clicks over time | Investigate substitution, decide upgrade or cut |
| High conversion, whatever the traffic volume | Small audience, disproportionate revenue | Protect aggressively |
| Strong topic, no unique asset yet | Could be owned, but hasn’t been differentiated | Build a moat |
| Topic with a clear recurring need | Users keep coming back to the same question | Turn it into a tool or tracker |
| Topic where real research is possible | Nobody’s published first-party data on it | Run the research once, let it compound |
“Traffic is down” isn’t a diagnosis on its own. A commodity page losing clicks and a high-conversion original-research page losing clicks need completely different responses, and lumping them together usually means panic-rewriting the wrong pages or ignoring the ones that matter.
Metrics that tell you more than total sessions
Total organic sessions is the wrong headline number right now, because it can mask exactly the shift that matters: stable visibility with collapsing economic value. Worth tracking instead:
- Search click yield: organic clicks divided by impressions, per page or query cohort. If impressions hold while clicks fall, something in the results page is absorbing attention, though that isn’t automatically an AI Overview.
- Revenue per 1,000 organic impressions: organic revenue divided by Search Console impressions, times 1,000. Joining that data is messy, so treat it as directional.
- Visitor-to-owned-audience conversion: signups or account creations divided by organic landing sessions, showing whether search traffic is becoming an audience you control.
- Returning visitor share, the clearest signal that content ownership rather than ranking is doing the work.
- Branded search growth: not perfect alone, but a rising trend of people searching your name or a methodology you’ve built is real.
- AI referral traffic where identifiable, treated as a floor rather than a ceiling, since a lot of AI-influenced behavior shows up as unattributed direct traffic.
- AI visibility or citation tracking, a useful leading indicator, kept clearly separate from the metrics above rather than treated as equivalent.
- Content substitution exposure: roughly what share of traffic, ad impressions, or affiliate revenue runs through pages classified as high substitution risk, telling you what’s structurally coming, not just what’s already happened.
A 30-day audit you can actually run
Week 1, measure. Export top landing pages from Search Console (impressions, clicks, click-through rate, ranking) alongside conversions and revenue, compared against a prior period. Look for pages with flat or rising impressions but falling clicks, stable rankings with dropping clicks, and the small set of pages actually responsible for most revenue.
Week 2, classify. Run the pages that stood out through the substitution-risk framework. Sort into high, medium, low. Cross-reference against revenue, the pages that matter are high-risk pages that are also economically important, not every commodity page on the site.
Week 3, build a moat where it’s deserved. Pick the highest-value vulnerable pages and add something genuinely hard to replace: first-party testing, an original dataset, a small interactive tool, real photography, an expert interview, or original research. The bar is materially improving the page, not looking good in a screenshot.
Week 4, capture the audience. Add one specific, compelling reason to come back directly: an alert tied to something the reader actually cares about, a saved tool, an account, a genuinely useful recurring digest. Measure the conversion rate before declaring victory.
None of this fixes everything in a month. It turns a vague anxiety about traffic into a short list of specific pages and specific actions, which is the only version of this problem that’s actually solvable.
What I’d stop doing
Publishing generic informational articles with nothing in them an LLM couldn’t produce from a two-sentence prompt: always a volume bet on ad impressions, a worse bet now that its query shape is exactly what AI Overviews target most. Obsessing over exact-match long-tail pages, when fan-out means Google increasingly understands the broader question rather than literal phrasing. Treating word count as a proxy for value: a 3,000-word rewritten summary is commodity content with better padding. Believing llms.txt is a ranking lever, when Google has said directly, twice, that it isn’t. Assuming schema markup makes content “AI-friendly” on its own, when it helps machines parse structure but manufactures no original insight. Measuring success purely through ranking position, when position one on a query an overview fully answers can be worth less in visits than position three on a query that doesn’t trigger one. Chasing AI citations as the goal, for the reasons above. And scaling AI-generated content just because production got cheap: when everyone can produce commodity text at near-zero cost, the scarcity that gave it economic value disappears with it.
Two monetization models under real pressure
If your revenue depends on affiliate commissions, be direct about which parts are exposed. A “10 Best Coffee Makers” post built from manufacturer spec sheets and a skim of other people’s reviews, with no product actually tested, is close to the purest example of commodity content there is, and precisely the shape an AI summary can reproduce without sending anyone your way. The stronger version isn’t complicated, just more expensive to execute: actually own and test the products, publish original photography and comparative measurements instead of copied specs, track long-term durability with follow-up updates, and use a scoring methodology that’s yours. Purchase intent still tends to produce outbound clicks eventually, since someone has to transact somewhere, a structural advantage commercial-intent content keeps over purely informational content. That protects the retailer at the end of the funnel, not a listicle that never did original work to get the visitor there.
If your economics run on display ads (search leads to an article leads to an impression), you’re structurally exposed if fewer of those searches convert into a visit. There’s no workaround for that math. What’s available: increasing revenue per visitor rather than chasing raw volume, reducing dependence on commodity queries most likely to trigger an overview, building a repeat audience that doesn’t require a fresh search each time, and exploring direct sponsorships that don’t depend on ad-network RPMs. None of it is a guarantee, just a set of levers that move a site’s economics away from its most exposed dependency, which is the realistic goal here.
Diversifying discovery without renting a different landlord
“Diversify your traffic” is usually left vague enough to be useless. Concretely: YouTube, Pinterest for visually driven niches, active (not promotional) participation in Reddit or specialist forums, syndication partnerships, referral relationships with adjacent sites. The catch worth naming: most of those are still rented audiences, subject to an algorithm or a rule change you don’t control. Diversifying across several rented channels reduces dependency on any one, which is genuinely valuable, but it isn’t the same as owning a relationship. The strongest diversification always includes at least one channel where you hold the actual contact: an email list, an account system, an RSS subscriber, a community you host yourself.
Brand is the moat, not “branding”
There’s a real difference between a search for “best CSS grid tutorial” and a search for a specific named source people already trust for that topic. One is a topic query any competent page can compete for. The other is effectively a navigational query wearing an informational query’s clothes, largely immune to AI summarization because the user already knows whose answer they want. Getting from the first to the second isn’t a branding exercise (a new logo, a consistent palette). It’s accumulated trust: distinctive expertise applied consistently, a recognizable and specific voice, original research published under your own name, tools people bookmark and return to, a track record specific enough that “I want to know what this person thinks” becomes reasonable. None of that happens from one article. It compounds from being right, specifically and verifiably, often enough that people start seeking you out by name instead of by topic.
Some pages should be allowed to die
Not every declining article deserves a rescue mission. If a page generates negligible revenue, contains no unique insight, costs real time to keep updated, answers a question an AI system already satisfies completely, and creates no subscriber or customer relationship on the way through, the rational move is often to stop maintaining it, consolidate it into a stronger page, or redirect it, and put your attention where it compounds. This runs against the instinct to defend anything with your name on it that used to get traffic. But spreading limited time evenly across everything you’ve ever published, regardless of whether it still earns that time, is how a site ends up with a thousand mediocre pages and no standout ones. Allocate based on where a page is going, not how much traffic it used to have.
This is bigger than Google
It’s worth stepping back, because framing this as “how do I deal with Google’s newest results-page feature” undersells what’s happening. People increasingly get answers from Google AI Overviews, Google AI Mode, Bing’s Copilot, ChatGPT, Perplexity, and Gemini, and the broader shift underneath all of it is the same one: search used to mean “find me documents,” and it increasingly means “answer my question directly.” I wrote about some of the adjacent shifts (spatial interfaces, AI-assisted products, edge computing) in a broader look at where web development is actually heading, and this is the content-economics version of the same pattern: the tools mediating between people and information are getting more capable of answering directly, and websites that only exist to hold an answer feel that shift first.
I don’t think traditional search is dying, and I don’t think websites are becoming irrelevant; both claims get made constantly and neither survives contact with how much traffic Google still sends. The more useful framing is narrower: a website increasingly needs to be more than the container holding the answer. It needs to give someone a reason to actually visit it, not just a reason to have it read on their behalf.
[TIAGO: optional personal example here, e.g. a specific page on this site where you noticed impressions holding steady while clicks dropped]
When agents start doing the browsing
Keeping this short, because it’s genuinely early, but it’s a distinct challenge from everything above. The next shift isn’t just AI summarizing your content for a human, it’s AI agents interacting with your site directly on a human’s behalf: comparing options, checking availability, filling in forms, eventually completing purchases. Google’s Chrome team has been shipping “auto browse” capabilities in Gemini that handle multistep tasks like researching flights or managing a subscription, pausing before an actual purchase, alongside an open commerce protocol co-developed with Shopify, Etsy, and others so agents can complete transactions on supported sites.
Google’s guidance for building agent-friendly websites treats agents as a distinct kind of visitor, one that reads your DOM and accessibility tree rather than your marketing copy, and that breaks on sites built around hover-dependent menus and unstable layouts. The reassuring part of that guidance is its own conclusion: “everything we suggest to make a site ‘agent-ready’ also makes sites better for humans,” meaning semantic HTML and stable layouts, not a separate AI-specific build. The distinction worth holding onto isn’t building for humans versus building for agents. It’s building a site that works correctly regardless of who, or what, is operating it, closer to accessibility discipline than a new category of SEO work. Worth watching, not worth rushing a redesign for.
What a defensible content site looks like from here
AI answer engines probably do erode part of the economic value that used to sit in commodity web content, and I don’t think that’s a temporary dip that reverses once the novelty wears off. The underlying capability, compressing widely available information into an accurate paragraph, keeps getting better, and there’s no version of “wait it out” that bets against that trend sensibly.
The response isn’t finding a clever way to get more commodity content linked from AI answers. It’s building a site whose value doesn’t live entirely in paragraphs a machine could already write: original data nobody else has, tools that only work when someone actually uses them, first-hand experience that shows in specifics rather than claims, a brand specific enough that people search for it by name, a direct relationship that doesn’t reset every time someone opens a new tab. Traditional informational content can, and probably should, stay part of how people find you in the first place. It just can’t be the entire business anymore, because the part that was easiest to replicate is exactly the part getting replicated.
If you’re trying to figure out where your own site stands on that spectrum, or want a second opinion on which pages are worth defending versus rebuilding, get in touch and I’ll give you a straight read on it.
FAQ
Does blocking AI Overviews with a nosnippet tag actually protect my traffic?
It can prevent that page’s content from being quoted inside an AI Overview: Google’s AI features documentation confirms existing nosnippet, data-nosnippet, and max-snippet directives apply there. But it doesn’t opt you out of ranking below other sites’ overviews, and the traffic you’re protecting has to outweigh whatever visibility you lose by blocking snippets entirely. For most pages, the substitution-risk framework above is a better use of your time than blanket blocking.
Will Google ever show me exactly how many clicks came from AI Overviews?
As of the generative AI performance reports that started rolling out in June 2026, Search Console shows impressions inside AI Overviews and AI Mode by URL, but not clicks, click-through rate, the underlying queries, or average position, and the reports are only reaching a subset of sites so far. Treat any AI-attributed click number from a third-party tool as an estimate, not a Search Console-verified figure.
Is it worth creating an llms.txt file for my site?
Not for Google Search specifically. Google has said directly, more than once, that these files don’t affect visibility or ranking in Search, including its generative AI features. If another AI product you actually care about reads and uses the file, there’s no harm keeping one, but don’t expect it to move Google traffic.
Do AI Overviews affect every niche the same way?
No, and the gap is large. Ahrefs’ trigger-rate research found overviews on over 40% of science and health queries but under 6% of shopping and real estate queries. If most of your content sits in a low-trigger category, the immediate urgency is lower, though the substitution-risk logic still applies to individual pages regardless of category.
Should I stop writing long-tail blog posts entirely?
No. Long-tail, specific content is still one of the more winnable positions for a smaller site, and query fan-out rewards being a strong, specific source on a narrow sub-topic. What’s worth stopping is producing long-tail pages that are purely commodity restatements with nothing original in them. Specificity plus originality holds up. Specificity alone doesn’t protect you the way it used to.
Final thought
None of this predicts that content websites stop mattering. It’s an argument that the ones worth building from here look less like a stack of paragraphs and more like a specific, accountable, sometimes interactive source that a machine can summarize but can’t actually replace.
