Which Content Formats Win the Most AI Citations for B2B SaaS?

Short answer: no single format wins. Citation share tracks query intent, not format quality. In our study of 490 cited URLs collected on 28 August 2026, comparison and alternatives pages took 56.7% of citations on head-to-head queries, pricing pages took 53.4% on cost queries, and listicles took 30.0% on category queries. The format that wins is the one shaped like the question being asked.

Almost every article on this topic hands you a ranked list: listicles first, then articles, then product pages. Those lists come from real datasets. They are also close to useless on their own, because the ranking flips depending on which questions were asked and which engine answered.

We ran the test ourselves to show why, and published the raw data so the coding can be checked.

What counts as an AI citation

An AI citation is a link an answer engine attaches to a claim in its response. In ChatGPT it appears as a small pill inside the text. In Google AI Overviews it appears as an inline chip and in the source panel on the right.

Three things people conflate, and the confusion explains most of the contradictory statistics in this category:

Metric	What it measures	Typical use
Share of citations	Of all cited links, what percentage were format X	"Listicles are 21.9% of citations"
Citation rate	Of all answers, what percentage contained at least one format X link	"Comparison content has a 95% citation rate in ChatGPT"
Share of a top-N set	Of the top 10 or 25 domains only, what percentage was X	"Reddit is 11.3% of ChatGPT's top 10"

These produce wildly different numbers from the same underlying data. A study reporting 21.9% and one reporting 95% can both be correct and are not comparable. Whenever you see a citation statistic without a stated denominator, treat it as unusable.

The short version, by query intent

If you only read one table, read this one. It is from our own run, both engines pooled.

The question your buyer asks	Citations sampled	Format that wins	Share
"Vanta vs Drata, which is better?"	120	Comparison or alternatives page	56.7%
"How much does helpdesk software cost per agent?"	133	Pricing page or pricing breakdown	53.4%
"Best alternatives to Zendesk"	73	Comparison or alternatives page	32.9%
"Best CRM for mid-market B2B"	100	Listicle or "best of" roundup	30.0%
"How do I choose a BI platform?"	64	Product or solution page	32.8%

Ask a head-to-head question and more than half of everything cited is a page whose only job is comparing two named products. Ask a pricing question and more than half of everything cited is a pricing page. The engines are not applying a general preference for one format. They are matching page shape to question shape.

How we ran the study

We wrote 40 buying-intent prompts across eight B2B SaaS categories: CRM, HRIS and payroll, contract lifecycle management, SOC 2 compliance automation, project management, helpdesk, marketing automation, and business intelligence.

Within each category we used the same five intents, because a software buyer does not ask one kind of question:

1. Category discovery: "best CRM software for mid-market B2B companies in 2026"
2. Head-to-head: "Vanta vs Drata for SOC 2 Type 2 compliance, which is better?"
3. Alternatives: "best alternatives to Zendesk for mid-market support teams"
4. Pricing: "how much does helpdesk software cost per agent per month?"
5. Selection: "how do I choose a BI and analytics platform for a SaaS company?"

Each prompt was run once in a fresh ChatGPT temporary chat, then once as a Google search where we recorded the sources inside the AI Overview. That is 80 live queries. Every cited URL was then classified by page type and by who owns the domain.

Result: 490 cited URLs, 151 from ChatGPT and 339 from Google AI Overviews, across 265 distinct domains.

Two counts worth knowing before the findings. Three of the 40 ChatGPT runs returned no citations at all, answering from model memory. Two of the 40 Google queries produced no AI Overview.

Finding 1: ChatGPT and Google are not the same game

Pooling engines hides the most actionable split in the data.

Format	ChatGPT (n=151)	Google AI Overview (n=339)
Product or solution page	37.7%	6.5%
Pricing page or pricing breakdown	20.5%	18.3%
Comparison or alternatives page	13.2%	23.9%
Product documentation or help centre	8.6%	0%
Review or analyst platform	6.0%	6.8%
Editorial article or guide	5.3%	12.7%
Listicle or "best of" roundup	5.3%	16.2%
Video (YouTube)	0%	6.2%
LinkedIn post or article	0%	3.2%
Trade press or media	2.0%	4.1%
Community forum (Reddit)	0%	0.9%

ChatGPT went to the vendor's own product and documentation pages. Google went outward to third-party comparisons, roundups, video and LinkedIn.

Two examples make it concrete. On "Zendesk vs Freshdesk for a B2B SaaS support team", ChatGPT cited three separate articles on support.zendesk.com, the vendor's own help centre, plus both vendors' pricing pages. On "Looker vs Power BI", every ChatGPT source was technical documentation: cloud.google.com, docs.cloud.google.com and learn.microsoft.com. Google, asked the same Looker question, cited YouTube, two LinkedIn posts and four independent blogs, and no vendor documentation at all.

Finding 2: product documentation is the format nobody optimises

Documentation was 8.6% of ChatGPT citations and 0% of Google's.

That is a larger ChatGPT share than listicles, editorial articles, or trade press. It is also the format most B2B SaaS companies treat as an engineering artefact rather than a marketing surface.

If your docs sit behind a login, render entirely in JavaScript, or block crawlers in robots.txt, you are invisible in a slice of ChatGPT that Google will never show you in a rank tracker. This is the cheapest gap in the dataset to close, because the content already exists. The crawler and rendering side of that is covered in full in our technical checklist post.

Finding 3: the listicle answer is a Google answer

The most repeated advice in this category is that listicles win AI citations. Our data says that is a Google finding being sold as a universal one.

Listicles were 16.2% of Google AI Overview citations and 5.3% of ChatGPT citations, a three-fold gap. In ChatGPT they ranked seventh of nine formats.

This collides head-on with the largest published page-type study. Wix Studio's AI Search Lab, using data from the Peec AI platform (both vendor-funded: Wix sells website software, Peec sells AI visibility tracking), published on 23 March 2026 an analysis of 75,000 AI answers and 1,056,727 citations. In its SaaS vertical, listicles were 35.37% of citations and comparison pages were 2.20% overall, with alternatives pages at 0.29% (Wix Studio AI Search Lab, Mar 23, 2026).

Our numbers are close to the inverse: comparison and alternatives pages at 20.6% of all citations, listicles at 12.9%.

The next section explains why, and it is the most useful thing in this article.

Why published citation studies disagree by 10x or more

Here is the same question answered by six credible datasets. The spread is not a measurement error. It is a design difference.

Study	Date	Sample	What was counted	Headline
Wix Studio / Peec AI (vendor-funded)	23 Mar 2026	75,000 answers, 1,056,727 citations, 3 engines	Every cited URL	SaaS listicles 35.37%, comparison 2.20%
HubSpot State of AEO plus Wix (vendor-funded)	upd 3 Jun 2026	"thousands of citation themes", Dec 2025 to Mar 2026	Answers containing a format	Comparison content 95% citation rate in ChatGPT
Omniscient Digital (vendor-funded)	20 Jan 2026, upd 25 Jun 2026	240 branded prompts, 23,387 sources	Every cited source	Reviews and social proof 57%
Foundation Inc with AirOps (vendor-funded)	12 May 2026, upd 30 Jul 2026	57.2M citations, 50 B2B brands, 7 verticals	Every citation, 7 verticals	Brand-owned content "a small fraction"
Citera (vendor-funded)	May 2026	~350,000 articles, 10,382 B2B SaaS keywords, 4 engines	Articles from Google's top 20	Brand-owned 29%, earned media 61%
Growfusely (vendor-funded)	28 Jul 2026	1,739 citations, 128 queries, 16 B2B software categories	One per unique domain per query	ChatGPT vendor pages 24%, review platforms 7.3%
Growth Memo, data via Profound (vendor-adjacent)	10 Aug 2026	3,177 SaaS prompts, ~35,000 URLs, ChatGPT, US	Every cited URL	Vendor domains 66.7% to 71.8%
NP Digital / Neil Patel	Aug 2026	"1,000 marketers surveyed"	Practitioner self-report	Listicles 21.87%, comparison 2.5%
This study	28 Aug 2026	40 prompts, 490 URLs, 2 engines	Every cited URL	Comparison 20.6%, listicles 12.9%, vendor-owned 57.1%

Four things drive the spread.

1. Prompt mix determines format mix. Eight of our 40 prompts were head-to-head and eight were alternatives, so 40% of our sample was made of exactly the questions comparison pages answer. Most studies do not publish their prompt distribution. Format share is not a property of an engine. It is a property of the question set you test. Any study reporting "listicles are X% of AI citations" without publishing its prompt mix is reporting its own prompt mix.

2. Branded and unbranded queries behave differently. Omniscient's 57% for reviews and social proof comes from 240 branded prompts, people asking about a named vendor. Ours were unbranded category and comparison queries. Both are real buyer behaviour and they produce different source mixes.

3. Denominators differ. HubSpot reports comparison content at a 95% citation rate in ChatGPT, meaning how often that format shows up among cited results. Wix reports 2.20% share of all citations. Neither is wrong. They are different questions.

4. Taxonomy differs. We counted a vendor's own /compare/vanta-vs-drata page as a comparison page. A study that files it as a product page will report a very different comparison share from identical raw data.

The practical takeaway: pick your formats from the queries your buyers actually type, not from someone else's aggregate. An aggregate built on a different prompt mix will point you at the wrong page.

Where our findings match the other studies, and where they don't

Agreement between independent samples matters more than any single headline number, so here is both sides of it.

Three findings that replicate across separate datasets

ChatGPT does not cite YouTube or Reddit for B2B software queries. Growfusely analysed 1,739 citations across 128 queries in 16 B2B software categories in July 2026 and found YouTube and Reddit citations "almost none" in ChatGPT, with all 47 YouTube citations and all 11 Reddit citations landing in Perplexity instead (Growfusely, Jul 28, 2026). We found exactly zero of each across 151 ChatGPT citations. Different month, different categories, different engine pair, same result.

Review platforms sit in the mid-single digits, not the double digits. Growfusely puts G2, Capterra and Gartner combined at 7.3%. We measured 6.5%. Growth Memo's SaaS study puts review platforms at 8.1% overall. Three samples, one range. That is a long way below how the category is usually sold.

Comparison-shaped questions pull comparison-shaped answers. Citera's May 2026 study of roughly 350,000 articles across 10,382 B2B SaaS keywords found AI Overviews triggered on 87% of comparison queries against 72% of best-of queries (Citera, May 2026). That is a trigger rate rather than a format share, but it points the same way as our finding that head-to-head queries are the densest citation opportunity in the funnel.

Two disagreements, and what actually causes them

Citera reports brand-owned content at 29%. We measured 57.1%. The gap is unit of analysis, not measurement error. Citera's universe is articles, harvested from Google's top 20 results per keyword. Pricing pages, product pages, /compare/ pages and documentation are largely outside that universe by construction, and those four formats are 65.5% of our ChatGPT citations. Citera is answering "among articles, how many cited ones are brand-owned", which is a genuinely different question from "of everything AI cites, how much is the vendor's own site". Their own write-up is admirably candid that format distribution in their sample was "too skewed to draw conclusions".

Growfusely reports ChatGPT vendor pages at 24%. We measured 78.8%. We cannot fully reconcile this one. Growfusely counts one citation per unique domain per query, which compresses the repeat vendor citations that inflate our number, and it reports "independent content plus vendor pages" as a combined 79% bucket, so its narrower "vendor web pages" line may exclude vendor blogs, docs and pricing that we count as vendor-owned. Directionally both studies agree ChatGPT leans on vendor properties far more than Perplexity does. On magnitude, treat the range as unsettled.

One number on this topic that we could not verify

Neil Patel's page on this query reports a content-type distribution attributed to NP Digital, with the disclosure "Data from 1,000 marketers surveyed" and a caveat that it "reflects practitioner-reported citation distribution rather than a controlled analysis of actual AI citation behavior".

Those figures are, to two decimal places, the citation-log distribution published five months earlier by Wix Studio's AI Search Lab from 1,056,727 measured citations.

Content type	Wix Studio (measured citations, Mar 2026)	NP Digital (stated as marketer survey, Aug 2026)
Listicles	21.88%	21.87%
Articles	16.68%	16.68%
Product pages	13.66%	13.66%
Category / hub pages	11.25%	11.24%
Other	9.92%	9.92%
Discussion / community	7.52%	7.53%
How-to guides	6.21%	6.22%
Homepage 5.26% + Profile 5.12%	10.38% combined	10.37%
Comparison 2.20% + Alternative 0.29%	2.49% combined	2.5%

Nine categories align within 0.01 percentage points, and the two pairs Wix reports separately sum to the merged figures. Wix and Peec AI are not credited anywhere on the page.

We are stating the arithmetic, not a motive. A survey of 1,000 marketers reproducing a citation log to two decimals is not a result a survey can produce, so one of the two labels on that page is wrong. Check it yourself before citing either version.

This is the practical reason to read methodology before headline: the same nine numbers appear on this SERP as measured citation data and as practitioner opinion, and only one of those can be true.

Finding 4: most B2B SaaS citations point at the vendor's own site

We classified each domain as vendor-owned, independent publisher, or platform. Small vendors we did not recognise fell into "independent", so the vendor-owned share below is a floor rather than a point estimate.

• All 490 citations: 57.1% vendor-owned, 24.5% independent publisher, 18.4% platform
• ChatGPT only: 78.8% vendor-owned, 11.9% independent, 9.3% platform
• Google AI Overview only: 47.5% vendor-owned, 30.1% independent, 22.4% platform

Two large published datasets agree on direction and disagree on magnitude.

Profound, which sells AI visibility software, analysed 11.84 billion citations collected between 16 April and 16 July 2026 across eight engines, covering 3.02 million domains. It found 57% of citations globally come from brand sites, with ChatGPT at 47%. It also found SaaS has the lowest reliance on earned media of any industry, at 11.4% median (Profound, Jul 30, 2026).

Kevin Indig's Growth Memo, using data captured in Profound, published on 10 August 2026 a study of roughly 35,000 citation URLs from ChatGPT, US only, December 2025, across 3,177 SaaS-vendor-related prompts. Vendor domains held 66.7% to 71.8% of citations at every buyer journey stage (Growth Memo, Aug 10, 2026).

One large dataset points the other way. Foundation Inc, working with AirOps, analysed 57.2 million citations across 50 B2B brands and seven verticals and reported that brand-owned content made up only a small fraction of citations, with Reddit at 28.0% on branded queries and 30.9% on unbranded (Foundation Inc, May 12, 2026, updated Jul 30, 2026).

We cannot reconcile these, and we are not going to pretend otherwise. The likeliest explanation is query class: Foundation's set spans verticals including productivity and DevOps where community and video platforms genuinely dominate, while ours and Growth Memo's are narrow, high-intent software-buying queries. If your buyers ask evaluation questions, the vendor-site finding is the one that applies to you.

What is clear either way: the widely circulated claim that "84% of AI citations come from third-party sites" has no traceable primary source and is contradicted by the two largest datasets we could verify.

Finding 5: Reddit, YouTube and LinkedIn are not what B2B SaaS buying citations look like

Across our 490 citations:

• Reddit: 3 citations, 0.6%. All three in Google, none in ChatGPT.
• YouTube: 21 citations, 4.3%. All 21 in Google, none in ChatGPT.
• LinkedIn: 11 citations, 2.2%. All 11 in Google, none in ChatGPT.

Growth Memo's SaaS-specific study reaches the same place from a much larger sample: Wikipedia, Reddit and LinkedIn account for 99% of UGC citations, and everything else, YouTube included, splits the remaining 1%.

The "Reddit dominates AI citations" claim is also unstable at source. Published figures for Reddit's share of ChatGPT citations range from 16.7% (Ahrefs, US queries, July 2026 data, sample size not disclosed) to 0.5% (Semrush reporting Promptwatch data, 14 to 17 August 2026, an 86% drop from a steady 3.8%). Search Engine Land covered the same drop on 19 August 2026 and reported the caveat plainly: Promptwatch could not rule out a data collection issue, and OpenAI said ChatGPT still cites Reddit.

Part of that spread is real change over time. Part of it is three different denominators being quoted as one number. Anyone giving you a single Reddit figure without naming the denominator is guessing.

Finding 6: there is no shortlist of sites to get on

490 citations came from 265 distinct registrable domains. 189 of them, 71%, appeared exactly once.

The most-cited domains were youtube.com (21), hubspot.com including its blog and knowledge base (17), salesforce.com (14), g2.com (11), vanta.com (11), linkedin.com (11), zendesk.com (10) and gartner.com (8). After that it is a long tail of small agency blogs, niche comparison sites and one-page affiliate pages nobody has heard of.

Two consequences. A placement strategy built on "get mentioned on the ten sites AI trusts" is chasing a set that does not exist. And the tail is winnable, because a site with no brand equity at all can hold a citation slot if its page answers the question in the right shape.

Review platforms held 6.5% of citations overall, real but well below how the category is usually sold. Worth noting that G2's own analysis with Kevin Indig, published 23 October 2025 using Profound data across 30,000 citations and 500 categories, found review volume explains under 2% of citation variance (R² = 0.009) (G2, Oct 23, 2025). That finding runs against G2's commercial interest, which is a reason to take it seriously.

What the evidence says does not work

Three tactics sold hard in this category have evidence against them. All three appear as confident advice on pages currently ranking for this query.

Schema markup. Ahrefs published on 11 May 2026 a matched difference-in-differences study of 1,885 pages that added JSON-LD between August 2025 and March 2026, against 4,000 matched controls. Result: Google AI Overviews −4.6%, statistically significant; Google AI Mode +2.4% and ChatGPT +2.2%, both statistically indistinguishable from zero (Ahrefs, May 11, 2026). The authors are explicit that they studied pages already cited heavily, so this does not test whether schema helps an uncited page earn its first citation.

Google's own Search Central documentation, last updated 10 December 2025, says it directly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." (Google Search Central)

FAQ blocks. A preprint on arXiv from 26 April 2026, using 602 controlled prompts and 21,143 search-layer citations, measured Q&A-format pages at −5.74% mean influence versus non-Q&A pages, while comparisons scored +55.28% (arXiv preprint, Apr 26, 2026). We flag this as weak evidence: no institutional affiliation, not peer reviewed, and the authors state their findings are descriptive rather than causal. Treat it as a reason to test, not a reason to delete your FAQs.

Word-count targets. Claims of a "2,500 to 3,500 word sweet spot" for AI citations circulate widely on pages ranking for this topic. We could not find a primary source for any of them. Our own dataset includes cited pages ranging from single-screen pricing tables to long-form guides. Treat unsourced multipliers such as "3.2x for FAQ schema" or "347% for entity optimisation" as marketing until someone publishes the method.

What to build, in order

Based on the intent breakdown, assuming a B2B SaaS company starting from nothing:

1. A pricing page with real numbers on it. 19.0% of all citations, 53.4% of citations on cost queries. "Contact sales" is a citation you hand to a competitor.
2. One comparison page per serious named rival. 56.7% of head-to-head citations. Rippling, Gusto, Vanta, HubSpot and Asana all had their own /compare/ pages cited on head-to-head prompts in our run. A vendor's own comparison page is not too self-serving to get cited. It gets cited.
3. An alternatives page for the incumbent you displace. 32.9% of alternatives-query citations.
4. Product and solution pages that state plainly what the product does. 37.7% of ChatGPT citations, the single largest ChatGPT format, and the one most often neglected in favour of blog volume.
5. Public, crawlable product documentation. 8.6% of ChatGPT citations and the format almost nobody treats as marketing.
6. Listicles last, and mostly for Google. 16.2% of Google citations against 5.3% of ChatGPT.

A note on self-promotional "best X" listicles specifically: they can earn a citation, but on the engine where buyers do their comparison work the format barely registers, and being listed in your own roundup is not the same as being recommended.

How to run this test for your own category

The aggregate in this article is a starting point. The version that matters is the one built on your buyers' actual questions. This method takes an afternoon and needs no tooling budget.

1. Write 20 to 40 prompts in your category, spread deliberately across the five intents above. Use the phrasing your buyers use, taken from sales call notes and support tickets rather than a keyword tool.
2. Run each prompt once in a fresh ChatGPT temporary chat. Temporary chat ignores memory and custom instructions, so results are not personalised to you. Repeat the same prompts as Google searches and record the AI Overview sources.
3. Log every cited URL against the prompt, the intent and the engine. A spreadsheet is enough.
4. Classify each URL twice: by page type (pricing, comparison, listicle, product, documentation, editorial) and by owner (you, a competitor, an independent publisher, a platform).
5. Read the gaps, not the totals. The useful output is not "listicles are X%". It is "on eleven of our head-to-head prompts, a competitor's comparison page was cited and we have no equivalent page".
6. Re-run monthly, not weekly. Track the format pattern rather than the URL list, for the reason in the next section.

If you would rather not run it by hand, that measurement layer is most of what an AI visibility tool or a program is actually selling.

How long do AI citations last?

Not long on ChatGPT, and this changes what is worth measuring.

SISTRIX, an SEO software vendor, published on 1 May 2026 the most methodologically transparent volatility study we found: 82,619 qualified prompts and 1,548,213 snapshots across six countries and three platforms, weekly from 17 December 2025 to 8 April 2026.

Weekly drift in the set of cited domains: Google AI Overviews 5%, Google AI Mode 56%, ChatGPT Search 74%. At URL level, ChatGPT drift rises to 85% per week. About 86% of prompts keep a stable core of a few domains while the rest rotates (SISTRIX, May 1, 2026).

Read against our data, this says something uncomfortable and useful. The individual URLs we recorded on 28 August 2026 will largely be gone from ChatGPT within a fortnight. The format pattern is far more durable than the domain list, which is exactly why building for format beats chasing placements.

One more piece of context on the size of the prize. Profound found on 3 February 2026, from roughly 700,000 US ChatGPT conversations sampled October to December 2025, that about 18% of conversations trigger at least one web search (Profound, Feb 3, 2026). The other 82% are answered from model memory, with no citations available to win.

Limits of this study

Stated plainly, because the numbers above are only as good as these caveats.

• 40 prompts is a small sample. Treat the intent-to-format pattern as the finding and the specific percentages as indicative.
• One run per prompt. Given SISTRIX's drift figures, a rerun next week would return a different domain list.
• Indian IP address, ChatGPT temporary chat on a logged-in account. Temporary chat ignores memory and custom instructions, but localisation still showed in vendor URLs.
• ChatGPT's hidden "+1" sources are excluded, so its citation counts are a floor.
• Page-type coding involves judgement. The classified CSV is published so the coding can be argued with.
• Almost every external study cited here is vendor-funded. The independent evidence base on this topic is close to non-existent, which is itself the most important thing to know about the numbers being quoted at you.

Note on verification: the original dataset in this article was collected by hand on 28 August 2026 from the live ChatGPT and Google interfaces, one query at a time, from an Indian IP address using ChatGPT temporary chat mode. Three limits we could not control. ChatGPT's citation pills sometimes carry a "+1" badge hiding a second source the page never exposes, so our ChatGPT counts are a floor. Some vendor URLs resolved to localised paths, so a US session may return a slightly different mix. Three Google runs returned more sources than our capture recorded in one pass and are marked truncated in the raw file. Every external study quoted here was re-fetched from the publisher's own page before publication. Where a study is published by a company selling GEO or SEO software, we say so at the point of citation.

---

Sources

• Content Types Most Cited by LLMs — Wix Studio AI Search Lab, 23 March 2026 (vendor-funded: Wix and Peec AI)
• Where Do AI Citations Come From? — Profound, 30 July 2026 (vendor-funded) · ChatGPT Citation Sources — Profound, 3 February 2026 (vendor-funded)
• Community Signals Are AI's Largest Third-Party Source — Growth Memo, 10 August 2026 (data captured in Profound, a vendor)
• AI Citation Patterns of B2B SaaS Verticals — Foundation Inc, 12 May 2026, updated 30 July 2026 (vendor-funded, with AirOps)
• Which Content Types LLMs Cite Most — Omniscient Digital, 20 January 2026, updated 25 June 2026 (vendor-funded, branded prompts only)
• An Analysis of 350,000 B2B SaaS Articles — Citera, May 2026 (vendor-funded; unit of analysis is articles, not all cited URLs)
• AI Citation Analysis for B2B Software — Growfusely, 28 July 2026 (vendor-funded; ChatGPT and Perplexity, domain-deduplicated)
• Which On-Site Content Types Earn the Most AI Citations — NP Digital / Neil Patel, August 2026 (figures match Wix Studio's published citation distribution to two decimal places while being attributed to a marketer survey; see the verification section above)
• On-page Content Formats Answer Engines Favor — HubSpot, updated 3 June 2026 (vendor-funded)
• Most Cited Domains in ChatGPT — Ahrefs, updated 21 July 2026 (vendor-funded, sample size undisclosed) · Does Schema Markup Predict AI Citations? — Ahrefs, 11 May 2026 (vendor-funded)
• Reddit's Citations in ChatGPT Fall — Semrush, 26 August 2026 (vendor-funded) · Reddit's ChatGPT Search Citations Fell 86% — Search Engine Land, 19 August 2026
• AI Features and Your Website — Google Search Central, updated 10 December 2025
• AI Citation Drift — SISTRIX, 1 May 2026 (vendor-funded)
• Do More G2 Reviews Mean More AI Visibility? — G2, 23 October 2025 (vendor-funded, G2 is the beneficiary)
• Structural and content features in generative engine citation — arXiv, 26 April 2026 (preprint, not peer reviewed, authors unaffiliated)