THE CITATION LEDGER · UPDATED SEPTEMBER 9, 2026

Gemini and ChatGPT barely cite the same sources

Ask both engines the same local-business question and the answers may sound similar. The source lists are not.

We matched 1,500 queries across 50 metros and 10 service categories. The average overlap between Gemini’s and ChatGPT’s cited domains was 8.3%. When both engines named a recognizable business, they chose the same business first just 4.9% of the time.

Same question, different source system

Gemini still leans toward businesses’ own websites. ChatGPT leans toward directories.

That is the main difference. It is large enough that “AI visibility” is not one ranking problem. A source that appears repeatedly in one engine may barely register in the other.

These are shares of citation references, not shares of queries. The same query can produce several citations, and the engines do not return the same number: Gemini typically returns about eight to ten citations per query; ChatGPT averages about 1.9.

That matters for the overlap statistic. The 8.3% figure is raw Jaccard overlap between the two returned domain sets: the domains in both sets divided by the domains in either set, or intersection ÷ union. It is not adjusted for their different sizes.

The overlap also varied by vertical, from 3.9% for dentists to 13.7% for electricians. Even the high end remained low.

The disagreement reaches the business names

The domain comparison tells us which domains each engine cited. The recommendation comparison tells us what a user sees.

Both engines named at least one recognizable business on 783 of the 1,500 matched September queries. Within that subset:

• The same business appeared first 4.9% of the time.

• Average overlap across the complete business lists was 5.2%.

This is not a claim that either engine is right. We did not score recommendation quality. It shows that a single rank position does not describe visibility across engines.

The business-name subset also changed between waves. Both engines named a recognizable business on 1,099 matched queries in July, compared with 783 in September—a decline of about 29%.

Gemini accounted for most of that change. Its no-recognizable-business rate rose from 25.8% to 47.4%. ChatGPT moved from 0.4% to 0.9%. A new Gemini filter for imperative advice bullets explains about 31 of the 328 added no-name cases; with the filter disabled, the September rate is still 45.3%.

“No recognizable business” is an extraction result, not proof that the answer contained no business. Gemini names are parsed from bold text at the start of list items, and that parser recognizes businesses in roughly 70–75% of queries. It can miss recommendations written in prose. ChatGPT names come from a structured field, so the two extraction methods are not identical.

Reddit left both engines, five weeks apart

On August 20, we reran the identical 1,500-query ChatGPT design after reports that ChatGPT Search had reduced its use of Reddit. Reddit went from 41.7% of ChatGPT citations to zero. A same-day follow-up also found zero Reddit citations in 334 citation observations.

At first, a 40-query Gemini spot check appeared to show no matching change. That check was too small.

The full September Gemini wave found Reddit at 1.7%, down from 13.7% in July. The two engines moved in the same direction, but not at the same time.

That correction is important. A one-engine or one-day snapshot can make a temporary difference look permanent. The dated waves are the finding; the platform story is an inference.

What this means for measurement

Do not combine Gemini and ChatGPT citations into one undifferentiated visibility score. Their source mixes differ too much.

Track them separately:

• By engine. A directory-heavy ChatGPT result and a website-heavy Gemini result are different surfaces.

• By query. Aggregate percentages can hide substantial variation by metro, vertical, and wording.

• By wave. The Reddit change shows that source behavior can move quickly.

• As a distribution. A single answer is an observation, not a durable rank.

This study observed citations and named businesses. It did not test whether changing a listing, page, or directory profile causes either engine to cite a business. Citation presence is also not the same as impressions, clicks, calls, or revenue.

Browse the matched-query data

The full matched-query dataset is available through the Citation Ledger raw-data page. It contains the query-level records used for this comparison.

How the comparison was built

The query matrix contains 50 US metros, 10 local-service verticals, and three prompt templates per metro-and-vertical pair. The literal query text was matched across engines.

Gemini data came from live Gemini API calls with Google Search grounding. ChatGPT Search data came through DataForSEO’s consumer-facing ChatGPT collection endpoint. ChatGPT location was country-level United States; city and state were supplied in the prompt text. All 1,500 ChatGPT queries completed and matched a Gemini key.

Before comparison, leading www. prefixes were removed. Without that normalization, equivalent domains would be counted as different sources.

Gemini business names were extracted from the answer structure with the project’s documented parser. It recognizes names in roughly 70–75% of queries and can miss prose-formatted recommendations. ChatGPT business names came from DataForSEO’s structured brand_entities field. The engines therefore do not have identical business-extraction methods.