Google now routes 30% of AI Overview citation links through a redirect. Check your citation tracking

Most of the AI visibility numbers a CPG brand looks at are built the same way. Someone asks a set of questions, records the links each answer cites, and counts how often a site shows up. That method has two quiet assumptions: that the link points at the site it names, and that today's answer looks roughly like next week's.

A study published on November 5, 2026 by Hasmord, a company that sells an AI search platform for B2B sales teams, tested both assumptions by accident. It was mainly about B2B buying questions, which are not our beat. But it reran a set of 60 shopping questions from October, and it noticed something about Google's links that every brand tracking AI Overviews should know.

What changed in Google's links

In the November 5 run, Hasmord found that 951 of 3,138 AI Overview citation links, or 30%, did not link straight to the cited page. They pointed to a google.com redirect with an encrypted address. On October 22, in its earlier shopping study, none of 2,333 links did.

Each wrapped link still displayed the site's name, so a person reading the overview sees the same thing as before. The difference is in the data underneath. Hasmord says it matched the names back to the sites using the direct links it had, and typed the rest by hand. As the study puts it, "Tools that read citations by URL will see google.com instead of the cited site."

We have not found a public statement from Google on the change, and one study on one day cannot tell us how widespread it is or whether it will last. What it can tell you is what to check.

Show the data

Run	Share of links
Oct 22, 2026 (2,333 links)	0%
Nov 5, 2026 (3,138 links)	30%

Source: Hasmord, November 2026 study. Oct 22 was a shopping-question run; Nov 5 a B2B-question run. Different question sets, same engine and method.

One caveat on the chart: the two runs used different question sets, shopping in October and B2B in November. That makes it a before-and-after of Google's link format, not of any one category.

Why this matters for a brand's numbers

Imagine your agency sends a monthly report that counts how often your domain appears among AI Overview sources for 200 category questions. If Google wraps a third of the links and the tool only reads URLs, three things happen at once:

• your domain's count drops, though nothing about Google's choice of sources has changed;
• google.com's count rises, which some tools already filter out or treat as Google's own pages;
• any trend line that spans the change shows a fall that is really a change in measurement.

This is the same trap we described when Gemini's referral tags disappeared: a change in how the platform labels traffic looks like a change in traffic. The fix is the same kind of fix. Ask the vendor or the analyst one question: are AI Overview citations matched by the URL only, or also by the site name shown? If URL only, the numbers since early November need a footnote.

Babel42's study from the day before is a useful reminder that Google's own links were already a large part of these answers. It left Google's shopping cards and links back to Google search out of its counts, and noted that they appeared in 26% of AI Overviews and 23% of AI Mode answers. We wrote about the shopping side of that in our post on AI Overviews citing Google's product pages. A wrapped redirect is a different thing from a Google product page, but both push in the same direction: a growing share of what looks like "Google" in a citation report is not one kind of source.

The second finding: sources drift over weeks

The other result is quieter and, for planning, more important. A reader had pointed out that runs an hour apart mostly measure caching. So Hasmord asked its 60 shopping questions again on November 5, two weeks after October 22, and compared the sites cited.

Runs minutes apart (Nov 5)Runs two weeks apart

Show the data

Engine	Runs minutes apart (Nov 5)	Runs two weeks apart
Perplexity	98%	62%
Google AI Overviews	82%	57%
ChatGPT	53%	37%

Source: Hasmord, November 2026 study. 60 US shopping questions. Same-day runs were 10 to 20 minutes apart. Averaged over all pairs of runs.

Perplexity, which repeated 98% of its sources within minutes, kept only 62% after two weeks. Google AI Overviews went from 82% to 57%. ChatGPT, already the least repeatable within a day at 53%, kept 37% over two weeks.

For a brand, that means a single snapshot of "which sites AI cites for our category" is a weather report, not a map. The set of pages that matter is real, but the specific pages cited on a given day move. Carson Cobb found something similar for brands rather than sources in a hand test of ChatGPT: the top pick matched on 10 of 11 repeated questions, while about 1 in 8 of the other brands changed. Sources drift more than picks do.

What shopping answers cite, for context

The study's shopping comparison is also a compact picture of where AI answers to shopping questions get their material. These figures are from the October 22 run.

Table: Share of answers to 60 US shopping questions citing each source type, Oct 22, 2026

Engine	Brand's own site	Media and publishers	Forum or social	Video
Google AI Overviews	53%	84%	67%	83%
ChatGPT	57%	82%	0%	0%
Perplexity	50%	99%	10%	0%

Source: Hasmord, shopping comparison table in its November 2026 study. G2 and Capterra counted as publishers in that run.

The pattern matches what other studies have found for consumer products. Publishers and reviewers dominate. Google leans on video and forums in a way ChatGPT and Perplexity do not. A brand site is cited in about half of answers, but that is any brand's site, not yours. For the work that follows, tip 11 on getting into the sources AI answers cite is still the starting point.

A 20-minute check you can run yourself

You don't need a vendor to see whether the redirect affects your category. Pick ten questions a shopper would type into Google about your products, such as "best gluten free pasta for kids" or "fragrance free laundry detergent for babies." Search each one in a fresh browser window signed out of Google, from a US location, and open the AI Overview when it appears. For each cited source, hover or long-press the link and note two things: the site name shown and the address the link actually goes to.

Write the results in a sheet with one row per citation: question, date, site name, link address, and whether the address is the site itself or a google.com redirect. Ten questions usually produce somewhere between 50 and 150 citations, which is enough to see whether wrapping is common in your category. Repeat the same ten questions a week later and again two weeks later. The share of wrapped links tells you how much a URL-only report is missing, and the overlap between the weeks tells you how much of your category's source list is stable.

This is a small sample, and Google's results vary by person and place. That is fine. The point is not to produce a benchmark but to know whether the numbers in your monthly report can be trusted as they stand.

The limits

Hasmord is a vendor, the B2B part of the study is outside CPG, and the methods are modest: 60 questions, one market, a few runs, site types assigned by one person. Perplexity was queried through its API, which returned at most ten sources per answer, and ChatGPT through a scraper of the search interface. The study says API and scraper results can differ from what a signed-in user sees. None of that undermines the two findings we have focused on, which are about the format of Google's links and the stability of sources over time. Both are easy to recheck yourself.

What we would do

Check the redirect in your own tracking. Pull last month's AI Overview citation data and search the raw URLs for google.com redirect links. If they appear and are being counted as Google, ask for them to be matched by site name instead, and annotate any trend line that crosses early November.

Stop comparing single days. Run the same question set at least three times over two to four weeks and report the share of runs in which your site or brand appears. That is a number that means something. A single day is a sample, and as the drift chart shows, a small one.

Separate "named" from "cited." OpenAI's new Business Hub counts a ChatGPT mention for an online business only when ChatGPT references the business's website, so it is a citation measure too. The brand-named measure has to come from your own checks, whether that is a shared sheet or SearchPanel from XENA Intelligence. Keep both, and expect them to move differently.

Don't chase the churn. If the specific pages cited for your category turn over by a third or more every two weeks, optimizing for whichever article was cited last Tuesday is wasted effort. The durable work is being present, and described correctly, across the kinds of sources that keep reappearing: retailer listings, reviewers, publishers and, on Google, video.

We said a while ago that brands should stop tracking AI rank as if it were a search position. This study adds a footnote: track citations as a rate over time, and check that the thing you are counting is still the thing you think it is.