AI Citation Index December 2026: Which Sources ChatGPT, Perplexity and Gemini Actually Cite

This is the third run of Anobee's AI citation index. The September 9 pilot set the prompt and the counting rules. The September 14 repeat checked whether the result reproduced. This edition asks a narrower question. After three observations, which pages behave like durable reference points, and which appear once and vanish?

"December" is the scheduled edition label. However, every table below uses the actual observation date, September 18, 2026.

AI citation index results at a glance

Each product received the same instruction in a fresh conversation or search: Search the web for Anobee and cite the sources you use.

Adding the three product counts gives 18. However, some pages appeared in more than one answer. After removing those overlaps, the cross-product union was 14 distinct Anobee URLs, up from 11 on September 14 and 13 on September 9.

No page reached all three successful products this time. ChatGPT, Gemini and Perplexity had all cited the homepage on September 14. This time ChatGPT left it out. Four pages reached exactly two products: the homepage (Gemini and Perplexity), plus About, the Privacy Policy and the zero-click search guide (ChatGPT and Perplexity). The other ten pages appeared in one product only.

What changed since the September 14 run

The pass rate did not move. Almost everything under it did, though.

Altogether, that is 17 URL-level changes across three products in four days, behind a headline that reads "no change".

ChatGPT, the most stable product in the previous comparison, lost four of seven pages. Gemini contracted to a single source. Perplexity had replaced its entire set between the first two runs. It replaced almost all of it again, keeping only the homepage. Meanwhile, the one product with a perfectly stable record, Google AI Mode, is stable because it has never understood the question.

Four pages have now survived every run

Three observations are enough to separate pages that keep reappearing from pages that appear once. Consequently, a durable-page list can start here.

Everything else has gaps. For instance, ChatGPT cited the homepage in the first two runs, then dropped it in the third. Perplexity has not cited any single page three times; the homepage is its only page with two appearances. Of the 25 Anobee URLs that have appeared in any run, 15 have appeared exactly once.

Complete September 18 AI citation index dataset

The table below is the complete normalized first-party dataset for the third run.

Google AI Mode has no column because it returned no Anobee URL. A page counts once per product. We removed tracking parameters (ChatGPT and Gemini append utm_source to their links) and merged trailing-slash variants. A CSV covering all three runs and all 25 pages accompanies this edition.

What ChatGPT cited

ChatGPT returned six explicit first-party destinations across seven inline citation groups. Specifically:

• About
• the Privacy Policy
• the SEO hub
• the zero-click search strategy guide
• the AI citation-checking guide
• the November 2026 edition of this index

ChatGPT has now cited three of those pages, About, the SEO hub and the AI citation guide, in all three runs. The homepage did not make the list. That is the first run in which any successful product left out anobee.com's front page.

Two things in the answer are worth recording. ChatGPT opened by stating that it was identifying the correct entity rather than a similarly named brand, and it was. The wider sources panel behind the answer held 30 entries, including lookalike domains such as anewbee.com and anabee.org. However, none of them reached the inline citations. ChatGPT also found and summarised the September 14 edition of this AI citation index and quoted the prompt back. Then it flagged that the page title says "November 2026" while the test date is September 14, and called the mismatch a possible freshness and entity-ambiguity problem for retrieval systems. The product under measurement had, in effect, reviewed the measurement. The point comes up again below.

ChatGPT displays citation groups as a chip such as "Anobee +1", which hides two destinations behind one marker. Only inspectable destination URLs entered the dataset. The counting unit is the URL, not the number of chips. OpenAI's help documentation says responses that use web search may include citations, that a Sources control shows cited sources and other relevant links, and that citations can be incomplete or incorrect and should be opened and checked [1].

What Gemini cited

Gemini cited one first-party page: the homepage. Additionally, it listed one other source, a Business of Apps profile of Amobee, and used it for an explicit disambiguation note that Anobee is distinct from Amobee, the ad-tech company.

That is a sharp contraction from the four pages of September 14. About, Online Business and Start Here all disappeared. Still, what Gemini did with its single source matters more than the count. Its answer described Anobee's six-step A.N.O.B.E.E. framework step by step, and every step matched the live homepage. One source, used well, produced a specific and accurate answer.

The Amobee note is the first time any product in this series has volunteered the disambiguation unprompted. The brand-name collision with Amobee is a known risk for Anobee. This run shows Gemini handling it correctly, at least once.

Google's Gemini help documentation explains that source and related-content links can appear inline or in a source panel [2]. This index recorded the destination URLs attached to the completed answer's source display. It did not add later double-check suggestions.

What Perplexity cited

Perplexity's source panel held 15 entries, and 11 of them resolved to anobee.com: the homepage, About, Contact, the Privacy Policy, the author page, the zero-click guide, the entity optimization guide, the llms.txt guide, both Rank Math guides and the Nepal earning guide. The answer text also linked five of those eleven inline: the homepage, the zero-click guide, the entity optimization guide, the llms.txt guide and the Nepal earning guide.

Compared with September 14, Perplexity kept only the homepage. Everything else in the set is new. Resources, the ChatGPT citation guide and the AI citation guide all dropped out. Ten pages entered. So, for the second run in a row, Perplexity's Anobee source set turned over almost completely. Its description of the brand stayed accurate throughout: an independent publication run by Bibek Thapa from Nepal, with the correct contact address.

The panel also carried noise. For example, one entry was the author's separate personal site. We logged it as an external source and excluded it from the anobee.com count. Three were unrelated: a Yahoo article about power outages, a list of eight-letter words beginning with N, and a Reddit thread about watches. None of them appeared in the answer text. As in the previous run, we recorded these as source-set noise rather than entity collisions, because the answer did not confuse Anobee with another organisation.

Why Google AI Mode recorded zero for the third time

Google AI Mode again read the fixed wording as a request for citation-generator tools. It returned "10 sites" under the heading "Here are top web results for exploring this topic": Scribbr, Cite This For Me, Citation Machine, Chegg, MyBib, EasyBib, Sourcely, two YouTube videos about APA style and a library guide. It produced no narrative answer. Instead, it listed tools. Nor did it name Anobee or link to any anobee.com page.

The zero stays in the dataset. Rewording the prompt after a poor result would break the comparison. Besides, three identical failures are themselves a finding: the branded prompt is reproducible and, for this product, reproducibly misread. Google describes AI Mode as breaking a question into subtopics and searching for each, and notes that it does not always get the answer right [3]. Here the subtopic it chose was "how to cite a website", and it answered that question competently.

A separate discovery panel with clearer, non-branded questions remains the right way to measure AI Mode. If added, it will appear as its own panel, never blended into this baseline.

What three runs of the AI citation index tell publishers

Retention is the metric, not presence

After three runs of this AI citation index, the presence metric has not moved once: 3 of 4, every time. The retention metric has moved every time. ChatGPT retained six of six pages between runs one and two, then three of seven between runs two and three. Gemini retained two of three, then one of four. Perplexity retained zero of seven, then one of four.

Thus a log that records only "cited: yes" would show a flat line and miss all of it. A log that records destination URLs shows which pages behave like reference points and which the products sample almost at random. The AI citation-checking method describes how to record prompts, destination URLs and answer accuracy so the second kind of log exists.

Four durable pages are a review list, not a ranking factor

The same product has selected About, the SEO hub, the AI citation guide and the homepage three times each. Therefore, those are the pages whose facts, ownership details and definitions most need to stay correct, because they are the ones most likely to be quoted.

It does not mean anything on those pages caused the citation. Three observations of one prompt cannot separate a page's wording from its age, its internal links, its position in Google's index or the product's own retrieval habits. Nevertheless, a short review list is more useful than a long one. The generative engine optimization guide covers the broader work without treating one source selection as proof of a ranking factor.

Entity handling improved while source stability did not

Every successful product identified Anobee correctly in this run. Gemini disambiguated from Amobee unprompted, and Perplexity's answer text stayed clean despite noise in its panel. That is better entity handling than the pilot, where Perplexity mixed in two lookalike domains. The work described in entity optimization for AI search, consistent naming across the site, the About page and the author page, fits these results. Also, the results cannot prove that the work caused them.

Source stability went the other way. Correct identification and stable sourcing are, evidently, different properties. A product can improve at one and not at the other, and in this run all three did.

Edition labels that lag the data create their own ambiguity

ChatGPT's unprompted note about the "November 2026" title on a September 14 test is a fair criticism from an unexpected reviewer. Naming editions by scheduled month while reporting an earlier test date is a convention this series inherited from its first edition. It keeps the archive in publication order. However, it also hands a retrieval system two dates to reconcile. This edition keeps the convention for continuity and records the objection.

AI citation index methodology

The third run of the AI citation index used the prompt, classification rules and normalization rules established in the pilot and unchanged since.

Test conditions

The products do not offer identical account, personalization or search controls. Those differences are part of the record, and we do not treat the sessions as equivalent.

Counting and normalization rules

A product passes when its original response displays at least one inspectable link resolving to anobee.com. Conversely, a brand mention without a first-party link does not pass.

Each normalized page counts once per product. We remove tracking parameters, ignore fragments and merge trailing-slash variants. For ChatGPT, we read the destinations behind grouped citation chips from the conversation's citation metadata and cross-checked them against the on-screen hover cards. For Perplexity, the count uses the answer's source panel, as in the previous two runs, with inline-linked pages noted separately. The author's separate personal site counts as external. Unrelated domains do not count even when a product treats them as supporting material.

Exploratory follow-up prompts stayed outside the dataset. We did not rerun a failed result until it passed.

Comparison rules

A URL is retained when the same normalized destination appears in consecutive runs for the same product. An addition appears only in the later run. Similarly, a loss appears only in the earlier one. A page is durable when the same product has cited it in every run to date.

The 75% figure divides three products with at least one Anobee citation by four tested products. Therefore, it describes this run only.

Limitations of the December AI citation index

This edition has one prompt, three observations and intervals of five and four days. It cannot estimate:

• Anobee's visibility for non-branded questions
• citation frequency across any product's full query space
• results in other countries, languages, accounts or personalization states
• how often a response would repeat across many trials on the same day
• the reasons behind page selection or omission
• the traffic, conversion or authority value of a citation

Google's own guidance for generative AI features says the best practices for SEO remain relevant and that no special markup is required to appear in them [4]. Nothing in this index contradicts that. Similarly, nothing in it identifies which practice produced which citation.

Two limitations are specific to this run. First, we counted ChatGPT's inline citations from citation metadata rather than by clicking each chip, because chips group several destinations behind one marker. The metadata and the visible hover cards agreed wherever we checked both. Second, we discarded one ChatGPT attempt before the recorded run, because the prompt accidentally went into an existing conversation rather than a fresh one. The dataset comes from the fresh chat only.

We prepared this edition ahead of its scheduled publication date. Its observations occurred on September 18, 2026. Use that date, not the issue label, when comparing the data.

What happens in the next AI citation index edition

The next AI citation index keeps the fixed branded prompt, the normalization rules and the failure labels unchanged. With three runs inside ten days, the series now has a short-interval baseline. Accordingly, the next run should sit further apart so that month-over-month comparison becomes meaningful.

The retention tables continue, and each edition updates the durable-page list. A separate non-branded discovery panel, with fixed questions and fixed denominators, remains planned. It will sit alongside this baseline, not inside it.

Meanwhile, subscribe for the next AI Citation Index to see whether ChatGPT returns to the homepage, whether Gemini's single-source answer holds, and whether any page becomes durable for Perplexity.