A version of this piece first appeared on leandrooliva.com.

It's tempting to say AI search killed the top of the funnel, and the recent traffic charts certainly make a decent case, but "the clicks are gone" is ultimately the precursory read. Nonetheless, the organic-search front door that fed acquisition for two decades is closing, and a different one is being built in its place: a funnel where visibility means being the passage an AI extracts and cites, not the page a human clicks through to. The plumbing, the metrics, and the tactics are all being reorganized around that idea. This is the ground now going by GEO (generative engine optimization) and AEO (answer engine optimization), and the last few weeks produced a run of developments that, read together, sketch what the new funnel looks like and what it rewards.

Most importantly, everything that follows comes with a warning attached: this is a moving target. In four days in mid-August, Reddit's share of ChatGPT Search citations reportedly fell about 86%, a swing that's still not fully explained, and a reminder that visibility inside someone else's AI is rented, with terms the platform can change overnight.

Here's what has caught my attention, and what I'd do about each one.

Google is quietly rewriting the scoreboard

The clearest number in the pile comes from a Pew study cited by The Washington Post: across 68,000 real searches, people clicked a result 8% of the time when an AI summary appeared, versus 15% when it didn't, a 47% relative drop.

Meanwhile, Chartbeat data in the Reuters Institute's 2026 predictions report shows organic Google traffic already down 33% globally and 38% in the US year over year, and news executives now expect search referrals to fall by more than 40% within three years, with a fifth of them bracing for losses above 75%.

That's the backdrop for Google's new Search Console feature, which lets you track how your TikTok, YouTube, and X posts perform in search results. One way to read it is as a reward for building a presence across platforms. Dan Taylor's argument in Search Engine Journal, however, is that it's closer to a shield. If your clicks fall by a third, Google can now point to social impressions and tell you you're still "visible," quietly moving the definition of success from "traffic we sent you" to "attention we showed on your behalf." Anyone that has been presenting analytics reports or been in the room for such a discussion will understand the intent here.

But to power generative search, Google has to work out which sites, profiles, and people are the same entity, and that is a process called entity resolution. When you verify that your site, your X account, and your TikTok all belong to you, you're hand-delivering exactly that map and doing work Google would otherwise have to infer. Those same profiles double as a "proof of life" signal separating real brands from the AI spam sites anyone can now spin up at scale. That data is getting more valuable to Google precisely as it gets cheaper, because you're the one supplying it.

Taylor also surfaces an interesting point on post deprecation. Evidently, AI systems will happily quote an old social post as if it were current. He found LLMs returning a client's pricing from a student-only offer posted on X back in July 2022. Old posts don't just fade into irrelevance, they can potentially resurface as misinformation about your own brand.

The takeaway: treat the new metrics as context, not success in and of itself, and keep measuring what you own. Verify your profiles if it genuinely helps you, but do it knowing what you're handing over. And go audit your own back catalog of social posts for anything an AI could quote back to a customer as today's truth.

Your Search Console data now contains AI conversations

Suganthan Mohanadasan noticed something subtle. Google has started leaking fragments of AI Overviews and AI Mode prompts into the ordinary Search Console performance report. Sitting alongside your normal keywords you'll now find conversational artifacts like "yes go on," longer multi-part prompts, tracker-bot noise, and pasted agent instructions, all attributed to your pages as if someone had typed them into a search box.

He built a free classifier to label and separate them, and notes the results are an undercount, since Google anonymizes rare queries and conversation strings are rare by nature. But the strategic point is bigger than the tool; the query data a lot of SEO reporting still rests on is now partly a transcript of AI conversations, not searches. If you're inferring intent or demand from GSC exports, some of what you're reading is people talking to an AI about your topic in ways they'd never phrase in a search box.

The takeaway: before you read intent into your query report, separate the searches from the conversations. What's left on the conversation side is a rough but real look at how people actually prompt AI about your category, which is worth studying in its own right.

The best paid creative might already be in the room

This is the most encouraging story in the set, and for some the most directly actionable. The Washington Post ran two campaigns side by side on Meta for six months, targeting comparable audiences. One used traditional direct-response creative that led with a subscription offer. The other used video its own journalists had already made, and led with the story. The journalism-led video delivered more than 4x the engagement, and, more importantly, beat direct response on actual subscriber acquisition, not just upper-funnel clicks. An earlier phase of the same strategy had already shown article-led promotion running 24% more efficient on cost-per-click and lifting conversion 31% over organic alone.

The why is the useful part. The Post's team argues that audiences, younger ones especially, don't have short attention so much as selective attention. They scroll past anything that reads like an ad and stop for something that feels real. A journalist's face and a trusted byline is a credibility signal no AI-generated creative can fake, and that gap widens every month as feeds fill up with synthetic content. There's even a structural tailwind: Meta's late-2025 algorithm update surfaces roughly 50% more Reels from creators who posted that day, so native, recent, human video gets a visibility edge that polished ad creative doesn't.

The deepest reframe is organizational. Editorial usually gets treated as a cost center that marketing pays for. The Post started treating it as its best-performing creative asset. Same content, different job. One strong piece of reporting can carry a paid campaign, a newsletter, and a media pitch at once, and the raw material already exists, sitting unused in the newsroom and the video library.

The takeaway: when performance dips, the instinct is to commission more creative. The higher-leverage move, when possible, is to look at what your team already publishes and put your single best real story to work across every channel, led by the story rather than the offer. You don't have to be a newspaper for this to apply, you need one genuinely good thing worth watching.

AI doesn't judge your page, it judges your paragraphs

Peec AI's breakdown of rerankers is the most technical piece here and the one I'd suggest practitioners take into account. The core reframe is that AI search is not a single decision: Discovery, passage selection, source selection, and answer generation are separate stages, and a failure at any one of them is invisible if you only track a single "visibility" score. So before you touch your content, you diagnose which stage failed. If the page was never retrieved, rewriting a paragraph can't fix it. If it was retrieved but never chosen, the passage is the problem, not the page.

The second reframe follows from the first: your page isn't judged as a URL, it's judged as passages. A strong page can contain a weak passage, and a mediocre page can hold the single cleanest extractable answer, which is exactly why page-level SEO metrics fail to explain so many citation wins and losses. In a test across 12 open reranking models, a descriptive product paragraph scored essentially zero for the query "best AEO tools," while a direct-answer rewrite of the same information scored above 99% on every single model. The models disagree on the exact numbers and agree completely on the direction.

Two practical rules come out of that. First, the cheapest high-impact edit is one direct-answer sentence in the first couple of lines of each target section, shaped to what the query wants: a shortlist for "best X," a definition for "what is X," ordered steps for "how to." Peec calls this intent-to-answer-shape alignment, and it's the real reason listicles win so often. Not because AI has a list bias, but because a good "best X" list packs named entities and extractable answers to many related queries into one place. Second, never trust one model's score. One model is an opinion; five models from three different architectures agreeing is a diagnosis. And when you report, split retrieval exposure, passage strength, and citation rate apart, because a single blended number hides the stage that's actually failing.

They also puncture a myth that's spreading in GEO circles: you cannot write for the chunker. Every system looks to split text differently, so matching imagined chunk boundaries is optimizing for a target you can't see. What actually helps is local self-containedness. Write so that any few-hundred-word window of your page makes sense read on its own, name your subject explicitly instead of leaning on "it" and "this tool," and keep each answer next to its evidence.

The takeaway: stop optimizing pages and start optimizing passages. Take your ten most important queries, find the exact section meant to answer each one, put a direct answer in its first two lines, and make that section stand on its own.

You probably don't need a content deal with OpenAI

A widely shared claim held that ChatGPT's in-house search index (nicknamed "Labrador") was effectively a closed, licensed tier reserved for big publishers like Reuters, the Guardian, and the Wall Street Journal, with the implication that without an OpenAI deal you simply weren't in the game. It didn't hold up when RESONEO ran the largest test yet, 1,249 ChatGPT answers across account types and countries, and found no difference in how licensed and unlicensed sites were served.

The reassuring read is that the playing field is more level than the panic suggested, but the catch is that level doesn't mean automatic. Labrador can only use what it can cleanly crawl and extract, which throws the emphasis back onto fundamentals. Serve static HTML so your content is readable without a browser rendering it first. Give important pages a clear, descriptive H1 that says exactly what the page covers. Answer the question in roughly the first 200 characters, because that's the part that actually gets stored. And build third-party signals through Reddit mentions, review sites, and the listicles that tell AI your brand is worth recommending. Just hold those borrowed signals loosely: Reddit's own citation share is exactly what lurched in mid-August, and a channel you don't control can be reweighted without notice.

The takeaway: you almost certainly don't need a content licensing deal. You do need pages a machine can read and a reputation it can find corroborated somewhere other than your own site.

The common thread

Every one of these is a symptom of the same rebuild. The old funnel ran on a ranked URL that earned a click, measured as traffic, which was then further segmented out. The new one runs on a passage that gets extracted and cited, and the definition of success is being rewritten by the platforms that benefit most from you accepting the new one. Those same platforms can reshuffle their favored sources overnight, as Reddit's whiplash just showed, which is why any visibility you rent on them is a tactic, never a foundation.

So the job splits in two, and you need both halves. First, become genuinely legible to machines: answer-first, self-contained, structurally clean, technically crawlable. That's table stakes now, and most of the reranker and Labrador findings are really about clearing that bar. Second, and this is the part the Post found, is build something a model can't hand back to the user for free. Original reporting, proprietary data, a trusted human voice, a direct relationship with an audience you actually own. Retrieval can only choose among the sources that exist, so the durable play is to be the source worth citing and to own the channel where your people already are.

As it stands today, the teams that treat AI search as a measurement problem may spend the next year watching a dashboard reassure them while KPIs quietly indicate otherwise. The ones who treat AI search as a content and trust problem will end up as the sources everything else gets built on, albeit with a watchful eye over their own tactics.