Methodology

Engines and exact versions

A study that does not name the model version it ran against is not reproducible, because the answer changes when the model does.

Engine	Model version
Claude	claude-sonnet-5
Gemini	gemini-2.5-pro
ChatGPT	gpt-5.5-2026-04-23
Google AI Overview	Google AI Overview

Run window, UTC: 2026-08-11T12:21:29.636790+00:00 to 2026-08-13T11:38:35.869239+00:00.

Method

The unit of analysis is one row per (company, prompt, engine). Each of the 50 companies in the Ahrefs cohort was assigned a category by the study team, for example, this is an expense-management tool, or this is an AI video generator, and then asked about using prompts phrased the way a real buyer would type them, generic category questions rather than questions that name the company. This produced 200 distinct prompts run across 4 engines: ChatGPT, Claude, Gemini, and Google AI Overview, for a total of 800 rows.

The design choice worth explaining is that every company is judged only on the question its own buyers would ask. A payroll company is asked about payroll software, not about video generation. This avoids a common failure mode in AI-visibility studies, where a company is marked invisible simply because it was asked about the wrong thing. It also means each company's result is comparable to the others in kind, if not in raw difficulty, since categories vary in how crowded they are.

Each engine was run once per prompt per company, with model versions pinned: 1 version of ChatGPT (gpt-5.5-2026-04-23), 1 version of Claude (claude-sonnet-5), 1 version of Gemini (gemini-2.5-pro), and Google AI Overview as it rendered live. The run happened over a window of 2 timestamps in August 2026, not a single instant, which matters because these systems are not static from one day to the next and a single-timestamp study risks capturing a transient state rather than a durable one.

Of 800 rows, 773 returned a usable answer, 8 rows were cases where Google did not show an AI Overview at all for that prompt, and 19 could not be parsed into a clean mention or non-mention. The 773 usable answers are the basis for every percentage in this study. Readers should treat the 19 unparsed rows as a small source of noise, not as hidden non-mentions or hidden mentions.

Limitations we volunteer

Written by us, before anyone else found them.

Terms used in this study

AI Overview

Google's AI-generated summary shown above traditional search results for some queries. It is treated as a separate engine in this study because it is generated differently from a conversational assistant and is not shown for every query.

appearance rate

The share of usable answers, across all prompts and companies in a cut, in which a given company or the cohort as a whole was named by the engine.

category_source=assigned

A flag on every row of the published dataset indicating that the category tested for that company was chosen by the study team rather than published by an external source. It is the single most consequential methodological choice in the study, since asking about the wrong category can make a real company look invisible.

cohort

The fixed group of fifty companies under study, taken unmodified from Ahrefs' published list of the fastest-growing SaaS companies by organic search growth.

displacement list

The set of companies an engine named instead of a given company, recorded for cases where the company being studied received zero mentions on that engine.

generic prompt

A prompt phrased as a category question, for example asking about expense management software in general, rather than a prompt that names the company directly.

mention position

The ordinal position, first, second, third, and so on, at which a company appeared within an engine's list of named companies, for answers where a mention occurred at all.

organic search growth

Growth in unpaid, non-advertised search traffic or ranking, the metric Ahrefs used to construct the fastest-growing SaaS list this study's cohort is drawn from.

unusable or not-extracted row

A row in the dataset where the engine's answer could not be reliably parsed into a clear mention or non-mention, and which is excluded from appearance-rate calculations rather than counted either way.

zero visibility

The status assigned to a company that received no mentions on any of the four engines tested, across every prompt asked about its assigned category, the strictest form of absence measured in this study.

References

Sources this study reads against. Every link was fetched and confirmed reachable at publication.

1. GEO: Generative Engine Optimization arXiv, 2023 Foundational paper formalizing generative engine optimization; introduces GEO-bench and reports visibility gains up to 40%, establishing that generative engines synthesize across sources rather than ranking pages.
2. Web search OpenAI API Documentation, 2026 Describes the Responses API web search tool that lets ChatGPT ground answers with sourced citations, relevant to how our ChatGPT prompts could retrieve current information.
3. The Most-Cited Domains in AI: A 3-Month Study Semrush, 2026 Independent study of more than 230,000 prompts over thirteen weeks finding Reddit and LinkedIn among the top five most-cited domains across ChatGPT, AI Mode, and Perplexity; used as an external comparison for our own top-cited-source table.
4. New G2 Research: Half of B2B Software Buyers Now Start Their Research With AI Chatbots PR Newswire / G2, 2026 Survey of 1,076 B2B decision-makers finding 69% chose a different vendor than planned based on AI chatbot guidance and a third bought from a vendor previously unknown to them, establishing the commercial stakes of the absence our study measures.
5. AI visibility is a topic-level game: A study of 50,000 brands in ChatGPT Semrush, 2026 Study of 1,094 subject areas in ChatGPT finding only 21% of the most-cited domains in a category are also the most-mentioned brand, a fragmentation finding that parallels our own split between visibility and total absence.
6. GEO: Generative Engine Optimization ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024 Peer-reviewed KDD version of the GEO paper; the venue of record for the method our study's engine-comparison design descends from.
7. GEO: Generative Engine Optimization ResearchGate, 2026 Critical survey reviewing 45 GEO studies from Nov 2023 to July 2026, arguing the field's terminology and evidence standards remain heterogeneous, which frames why we specify our own method in full.
8. ChatGPT Search OpenAI Help Center, 2026 Vendor documentation confirming ChatGPT search answers may include inline, clickable citations, the mechanism our appearance measurements rely on existing at all.
9. AI search engines cite Reddit, YouTube, and LinkedIn most: Study Search Engine Land, 2026 Reports a Peec AI analysis of 30 million sources finding Reddit, YouTube, and LinkedIn are the most-cited domains in AI-generated answers, corroborating the composition of our own top_cited_sources list.
10. How AI tools shape the B2B buying process: A survey of 600+ US business professionals Semrush, 2026 Finds 71% of respondents use ChatGPT for product research and 61% use Google Gemini for the same, supporting our choice of engines as the ones buyers actually use.
11. Gartner Survey Finds Sixty-Nine Percent of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights Gartner, 2026 Survey of 645 B2B buyers finding an average of seven information sources used per purchase and 45% using generative AI mainly to gather vendor and product information, contextualizing why an engine's silence on a vendor matters.
12. The conference for marketers ready to win in 2026 (Top 10 Most-Cited Domains in AI Assistants) Ahrefs, 2026 Analysis of roughly 76.7 million AI Overviews, 957,000 ChatGPT prompts, and 953,500 Perplexity prompts finding Wikipedia is the most-cited domain across all three engines, cited by 16.3% of ChatGPT answers, 12.5% of Perplexity answers, and 8.4% of AI Overviews.
13. Google's Guide to Optimizing for Generative AI Features on Google Search Google Search Central, 2026 Google's own position that AI Overviews are rooted in core Search ranking and retrieval, used here to explain why classic SEO winners are not guaranteed AI-answer winners.

Data

The complete row-level dataset is published open and ungated under CC BY 4.0. Every number in this study can be recomputed from it.