Google's AI Overview Links Now Point Back at Google: The Two Anchors Caught on Sept 21 Reroute Search's Exit Into Its Own Surfaces

On 21 September 2026, SEO watcher Gagan Ghotra recorded a short clip on X: at the bottom of a Google AI Overview sat several blue anchors that looked exactly like ordinary web links. Clicking them took you to no website at all. They pushed you into AI Mode. The next day, 22 September, Search Engine Roundtable wrote the test up. In the same week, Google Discover began testing a "Dive deeper" button that opens an AI-generated topic card rather than a publisher's article. Google's two largest free referral exits were wired back into Google's own surfaces in one week.

This is not an isolated event. It is an eighteen-month straight line. AI Overviews now appear on roughly 48% to 60% of searches and reach over two billion people monthly. Websites pay for that reach: Google referral traffic to publishers fell about a third globally in a year, small sites at 1,000–10,000 daily page views lost roughly 60% over two years, and mid-size sites at 10,000–100,000 lost about 47%. Alphabet's own filings tell the other half of the story. In Q1 2026, Google Search & Other revenue hit $60.4 billion, up 19% year over year, while Google Network — the money paid out to external sites — fell 4%. Search revenue is growing; the share handed outward is shrinking. These anchor links simply upgrade the trend from "fewer clicks for you" to "the exit no longer points at you."

Compare the other platforms and the direction is obvious. ChatGPT and Perplexity still put citation links beside the answer, and Perplexity sells its source list as a feature. Bing keeps an explicit Web tab. Google went the other way: AI Mode layered over AI Overviews, AI cards layered over Discover, and Web Guide blending both interfaces — with the "Classic search" button caught this week not actually returning you to classic search. The category is not converging. It is splitting into "sources as product" and "sources as raw material," and Google picked the second. Worse for the narrative, research publicised by Jason Kint on 22 September argues AI Overviews improved neither satisfaction nor visit quality, and only reduced clicks to their sources.

So what should a Taiwanese SMB, a clinic, or a studio that builds websites do? The short version: your GA4 will break before your rankings do. What follows is how to measure this within 30 days, which three traffic sources to shore up now, and a DIY monitoring setup that costs nothing in SaaS subscriptions.

What Actually Happened: Two Exits, One Direction

Exit one: the AI Overview anchor link. Ghotra's capture shows links at the bottom of an AI Overview, visually identical to normal web links, that carry a follow-up question and drop the user straight into AI Mode. His comment was blunt: another tactic to push users from ordinary search results into AI Mode. It is a test, not a full rollout.

Exit two: Discover's "Dive deeper." Google's Robby Stein announced the experiment on X on 18 September. Tapping it shows a short topic overview labelled as AI-generated — with Google's own caveat that it can make mistakes — and only then offers links to publisher stories. Stein said the test starts on video content with multiple designs to follow. Barry Schwartz's read: if the generated summary already supplies the full context, why would anyone click?

A third thing the same week: Google Business Profile started emailing about spam review spikes. The subject line is "Protecting your Business Profile from spam reviews," telling owners the system detected and removed anomalous reviews. It looks unrelated, but it is the same logic. Google is moving judgement — who is credible, what counts as the answer — inside its own box, and you receive the outcome as a notification.

One more signal worth logging: public Gemini Notebook pages are being stuffed with spam and pushed into Google Search. When a platform's own generated surfaces have a quality-control problem, it has more reason to keep users inside its interface, because inside its interface it can switch things off.

What Each Reader Should Do Today

Brand owners and SMB operators

• Open GA4 today and check one number: what share of the last 90 days' sessions came from google / organic. Above 60% means the business is leaning on a channel that is closing.
• Move part of the budget from "write more blog posts" to owned audience: email, LINE OA, SMS. No algorithm can switch those off overnight.
• Google Business Profile is still the highest-value remaining exit for local businesses. Start reading those review-anomaly emails weekly rather than letting reviews vanish silently.

Marketers and SEO practitioners

• Replace rankings with impressions-to-clicks as your headline metric. Flat position with collapsing CTR is exactly what an AI Overview eating your slice looks like.
• In Search Console, cut a cohort of informational queries (containing how / what / why) and track it separately. That cohort dies first.
• Re-tier your content: rewrite informational long-form into citable paragraphs — one clear conclusion each, with a number and a date — and concentrate effort on transactional and local pages, which still keep their clicks.

Developers and agencies

• Add referrer classification: tag traffic arriving from google.com with AI Mode parameters separately, or it all disappears into undifferentiated organic.
• Upgrade schema from "implemented" to "maintained." Article, FAQPage, Product and LocalBusiness fields must track content changes; stale schema is worse than none.
• Open a service line: a quarterly AI visibility audit delivering a rewrite list of citable paragraphs.

Tool Comparison

What Nobody Will Tell You

• The "better experience" claim has counter-evidence. Google argues AI Overviews deliver a better experience and higher-quality clicks. Research publicised by Jason Kint on 22 September says satisfaction did not improve, visit quality did not improve, and only clicks to sources fell. This is the first head-on collision between the PR narrative and third-party measurement.
• It is a test, not policy — but tests never roll back. The anchor links appear on some queries only. Over the past eighteen months, every Google "test" became the default; the only variable was six weeks versus six months.
• AEO and GEO tools sell proxy metrics. There is no verifiable bridge between "mentioned N times in AI answers" and revenue. Until Google ships official attribution, every AI visibility score is a sampled estimate. Do not report it to your boss as a KPI.
• Rank tracking is drifting from reality. The "position 3" in your report and the screen on a user's phone are no longer the same object. Keeping rankings as an acceptance criterion makes both you and your client decide wrongly.

The No-Subscription SMB Alternative

Search Console API plus Looker Studio, two days of work, produces a CTR-collapse detector for informational queries. The logic is simple: query groups where impressions and position hold steady but CTR declines week over week are the slice the AI answer took.

# 1) Pull 90 days of query-level data (enable the Search Console API in GCP first)
python -m pip install google-api-python-client pandas

# 2) Query by query + date; rowLimit 25000, paginate to exhaustion, write CSV

# 3) Split out the informational cohort
import pandas as pd
df = pd.read_csv("gsc_90d.csv")
INFO = ("how ", "what ", "why ", "best ", "vs ", "guide")
df["bucket"] = df["query"].str.lower().apply(
    lambda q: "info" if any(k in q for k in INFO) else "commercial")

# 4) Weekly rollup: find flat position with falling CTR
w = df.groupby([pd.to_datetime(df.date).dt.to_period("W"), "bucket"]).agg(
        impressions=("impressions", "sum"),
        clicks=("clicks", "sum"),
        position=("position", "mean"))
w["ctr"] = w.clicks / w.impressions
print(w.round(4))   # stable position + 4 straight weeks of CTR decline = the AI slice

Wire it into Looker Studio with one rule: email an alert when informational CTR falls more than 15% over four consecutive weeks. That is the zero-subscription early-warning system. Treat the official Search Console performance report documentation as the baseline definition.

FAQ

If AI Overview links point into AI Mode, how much traffic will I lose?

It is a test and there is no official figure. For scale: Google referral traffic to publishers fell roughly a third globally over the past year, and small sites lost about 60% over two years. Sites weighted toward informational content are hit hardest; transactional and local pages hold up better.

Should I block AI crawlers?

For almost every SMB, no. Blocking does not return clicks; it removes you from AI answers while your competitors stay in them. Crawler blocking is a negotiating lever for large publishers, not a defence for a small business.

Should I still track rankings?

Track them, but demote them to a reference metric. Make Search Console impressions, clicks and CTR the headline numbers, because under AI interfaces the same nominal position maps to wildly different real estate.

Where should the money go right now?

Three places: owned audience (email and LINE OA), Google Business Profile review and information hygiene, and rewriting existing content into citable paragraphs. The first two pay off immediately; the third at least keeps your brand name inside AI answers.

My Take

The consensus line is "SEO isn't dead, it just became GEO." I disagree in a specific way: for SMBs, the word GEO is currently more useful for selling consulting than for making decisions. The reason is attribution. SEO became a business because Search Console supplied official impressions and clicks. GEO has no equivalent, so every AI visibility score is third-party sampling. Betting budget on an unverifiable metric is not meaningfully different from betting it on rankings — it just has a newer name.

The more useful framing: the real structural opportunity is not "how to get cited by AI" but "how to stop depending on any single platform's exit." For a studio like ScriptWalker that translates into a concrete service line — package owned-traffic infrastructure as a fixed deliverable: an email list system, segmented LINE OA push, on-site search plus content retargeting, and the zero-subscription CTR-collapse monitor. Together roughly NT$120,000–260,000. What the client buys is not rankings; it is "when Google closes another exit, I still have something that works." That is easier to sell, and far easier to deliver, than an AI visibility report nobody can interpret.

Sources

• Primary: Gagan Ghotra's demonstration on X (21 Sep 2026)
• Primary: Google's Robby Stein announcing the Discover "Dive deeper" experiment (18 Sep 2026)
• Primary: Google Search Console performance report documentation
• Third party: Search Engine Roundtable — AI Overview links pushing to AI Mode (22 Sep 2026)
• Third party: Search Engine Roundtable — Google Discover tests Dive Deeper
• Third party: Search Engine Roundtable daily recap (22 Sep 2026), incl. Jason Kint research and Business Profile emails
• Third party: Search Engine Journal — AI Overviews impact on publishers
• Third party: Advanced Web Ranking — Why tracked rankings no longer match Google