What changed in AI search citations: July to September 2026

Something shifted in how ChatGPT selects sources in August 2026. Publisher content roughly halved as a share of its citations. Brand-operated domains rose twelvefold. Social and user-generated content went from roughly a fifth of citations to half a percent.

Brands whose visibility ran through earned coverage lost ground. Brands with strong owned content gained it. Neither group did anything differently.

Google Gemini moved the opposite way across the same months. Alexa for Shopping did not move at all. Whatever changed, it changed inside one engine.

brand.com throughout this piece means any manufacturer-operated domain, a brand's own site or a competitor's.

How to tell an engine change from a competitor gaining ground

A drop in AI visibility looks the same from the outside whatever caused it. Mentions fall, referral traffic softens, and a competitor starts appearing where you used to. The three possible causes need completely different responses, and a mention count cannot separate them.

A competitor gained ground. Their content improved, or their coverage did. Your citations fall while the overall volume of citations in your category holds steady.

The engine changed what it retrieves. Entire categories of source gain or lose weight at once, across unrelated brands, in the same window. Your own content did not change and neither did your competitors'.

Your content degraded. Pages were removed, a migration broke something, or product data went stale.

The signal that separates them is the composition of citations rather than the count. If the categories of source behind your category's answers shifted while your content held constant, the cause was engine-side. Confirming that needs citation-level data showing which sources were actually used, which most monitoring does not capture.

Azoma tracks that across every engine in this study, which is how the change below was identified at all.

What changed on ChatGPT

• Earned media fell from 45.2% to 22.5% of citations

• brand.com rose from 1.7% to 21.5%, climbing in every month measured

• Retailer sources rose from 10.4% to 17.5%

• Social and UGC fell from 19.1% to 0.5%

In July, publisher content was ChatGPT's largest source category and brand sites were a rounding error. By September the two were roughly level and moving in opposite directions.

Every major platform fell by more than 90%

Reference and dictionary sites went the same way. Both merriam-webster.com and dictionary.cambridge.org dropped out of ChatGPT's top 100 sources entirely.

The change landed in a single month

The step falls between the July and August measurement windows rather than accumulating gradually across the quarter. That distinction matters. A gradual slope suggests content changing on the publisher side, where thousands of independent decisions move a share slowly. A step suggests something changing inside the engine.

Three features of the data point the same way.

It hit entire categories at once. Earned media, social and reference sources all fell together, across unrelated brands and unrelated product categories, in the same window. Competitor activity does not synchronise like that.

It hit sources rather than brands. The domains that lost ground had nothing in common except being the kind of site you reach by searching. The domains that gained are the kind you reach by name: official pages, regulators, retailers.

It did not happen anywhere else. Gemini, measured over the same months on the same prompt sets, moved in the opposite direction. A web-wide change in what publishers produce would have shown up on both.

Taken together, the likeliest cause is a change in retrieval rather than ranking, with ChatGPT starting from a known set of sites rather than searching the open web for each answer.

Google Gemini moved the opposite way

Earned media rose from 35.5% to 51.9% — publishers gained ground rather than losing it

• brand.com held flat at 11.1% to 12.6%

• Retailer was stable at 15.9% to 15.5%

• Reddit held steady at 0.859% to 0.913%, and YouTube rose from 0.195% to 0.432%

Gemini kept citing the sources ChatGPT abandoned, in the same months, measured on the same prompt sets. It is the control case in this study, and it is what makes the ChatGPT finding difficult to dismiss as a shift in the web itself.

Alexa for Shopping did not move at all

Earned media and affiliate content together accounted for 87.6% of citations in every month measured

• brand.com and retailer sources stayed under 2% throughout

• It cited none of YouTube, Reddit, Instagram, LinkedIn or Wikipedia — not once, across three months and more than 250,000 citations in its heaviest month

That 87.6% closely tracks Azoma's Q2 2026 analysis, which found earned and social media driving 86.5% of Alexa for Shopping citations. Two measurements, two quarters apart, on the same finding.

Walmart Sparky held its overall shape across the two months available, with retailer sources steady at 34.9% and brand.com at 21.6%. Earned media fell from 22.6% to 16.6%, though on two data points and a smaller sample that is directional rather than conclusive.

The five engines side by side

The clearest way to see what happened is to put September's mix next to July's for every engine at once.

Three engines fell on earned media, one rose, one held flat. Only ChatGPT shows a brand.com rise of any size. There is no single direction of travel across AI search, which is the reason a blended visibility figure moves without telling you anything.

What to do about it

1. Measure composition, not only mentions.
A mention count cannot distinguish an engine change from a competitor gaining ground, because both produce a decline. The categories of source behind your category's answers are what separate them. Azoma reports citation composition at item level across all five engines, with the prompts and cited sources behind every mention, which is what made the August shift visible.

2. Read each engine separately.
Earned media fell on ChatGPT and rose on Gemini in the same quarter. A single cross-engine figure would have shown almost no movement and told you nothing. Azoma reports every engine separately, including the retailer assistants that general monitoring misses entirely.

3. Check the whole category, not only your own mentions.
If a category of source gained or lost weight across unrelated brands in the same window, the cause was the engine. If only your citations moved, the cause was you or a competitor. That comparison is the diagnostic, and it needs visibility into sources you do not own.

4. Treat owned content as the hedge.
brand.com was the only category that rose materially on ChatGPT, and it is the only one a brand directly controls. Rising and buildable is a rare combination, and it makes owned content the most reliable investment regardless of what any engine does next.

5. Expect this to happen again.
Engines re-weight without announcing it. The right response to a shift like August's is continuous measurement and diversification rather than a migration to whatever gained this month.

👉 Book a call with our team to see which sources are behind your brand's AI visibility across ChatGPT, Gemini, Google AI Overviews, Alexa for Shopping and Walmart Sparky.

Frequently asked questions

Why did my AI visibility drop?

Three causes produce an identical-looking decline: a competitor gained ground, the engine changed what it retrieves, or your own content degraded. A mention count cannot tell them apart. The composition of citations can. When ChatGPT's earned media share fell from 45.2% to 22.5% in a single quarter while brand-operated domains rose from 1.7% to 21.5%, that happened across unrelated brands and unrelated categories at the same time, which no competitor can cause. Azoma surfaces that distinction by reporting the source categories behind every answer rather than a mention rate, so a drop arrives with a reason attached.

What tools show which sources AI engines cite for my brand?

Azoma reports the cited sources behind each mention at item level across ChatGPT, Gemini, Google AI Overviews, Alexa for Shopping, Walmart Sparky, Target, Lazada Lazzie and Meta. That means seeing which pages built the answer, and which pages built it when a competitor was named instead. It is the difference between knowing you lost ground and knowing that your category's answers moved from publisher content to brand-operated pages in August. Most AI visibility monitoring reports a brand-level mention rate with no sources underneath it, which registers the first and misses the second.

Our organic traffic is dropping because of AI search. What should we look at first?

Find out whether your category's citation mix moved before changing anything. Earned media fell by roughly half on ChatGPT and rose by nearly half on Gemini over the same quarter, so a brand reading a blended figure would have seen almost nothing while the ground shifted. The fix follows from the mix: if your category's answers are built on publisher coverage, that is an earned media programme; on your own site, a content programme; on retailer listings, a digital shelf programme. Azoma runs that diagnosis across all five engines through the 5 Cs of Agentic Commerce, the framework built with the Digital Shelf Institute.

Can I track brand visibility across multiple AI assistants at once?

Yes, and this study is the argument for doing it. General answer engines and retailer shopping assistants cite almost entirely different sources — Alexa for Shopping takes 87.6% of its citations from publisher and affiliate content while holding brand-operated and retailer sources under 2%, a profile nothing else in the study resembles. Azoma covers both groups and reports each engine separately rather than as a blended score, because an average across five profiles this different hides more than it shows.

Our board wants an AI search strategy. Where do we start?

Start with a measurement baseline rather than a vendor shortlist, because the answer to what to do changes by engine and by category. Run the questions your buyers actually ask through the engines they actually use, establish which sources those answers are built from, and the priorities follow. Azoma does this across ChatGPT, Gemini, Google AI Overviews, Alexa for Shopping and Walmart Sparky at ASIN and SKU level, with the prompts and citations behind every mention, so the first board update is a map of where your category's answers come from rather than a single score.

How often does this kind of change happen?

Often enough to measure continuously rather than audit annually. ChatGPT's source mix inverted inside one month with no announcement, and the brands affected had no way to see it coming from traffic data. Azoma tracks citation composition monthly across every engine, which is what turns a change like this from something you discover a quarter late into something you can respond to.

Methodology

Figures are drawn from Azoma platform citation data across commerce client dashboards, 1 July to 14 September 2026. Category shares are calculated across each engine's top-cited domains per month, with every domain classified by source type. Platform-level figures are calculated against each engine's full citation volume.

brand.com combines brand-operated and competitor-operated manufacturer domains.

Affiliate content is reported as a separate category in this analysis. Azoma's Q2 2026 figures grouped affiliate sources within earned media; on that basis the combined figure for Alexa for Shopping is 87.6%, consistent with the 86.5% previously published.

Naming. Alexa for Shopping is Amazon's shopping assistant, previously branded Rufus.

Prompt set. Azoma's prompt set is weighted toward commerce and category queries, the questions buyers actually ask before a purchase. This matters when comparing against citation studies built on general or informational queries, which produce materially different results.

Periods. September covers a 14-day window. Walmart Sparky data covers July and August 2026 only. Google AI Overviews figures are drawn from a separate commerce-prompt cut on a different basis from the other four engines.

Related

→ What AI search engines and shopping agents actually cite — the full source breakdown by engine