Research · Evidence review
The State of AI Search Visibility 2026
How AI systems retrieve, cite and recommend brands, what that does to demand, and how to measure it without fooling yourself. For marketers, SEO practitioners and agencies.
A brand can be described accurately by an AI assistant and still never be recommended. A page can be retrieved and never cited. A recommendation can create demand that never shows up as an AI referral. These are different outcomes, and the research published in 2026 shows why treating them as one “ranking” gives a misleading picture of AI visibility.
This review brings together the strongest evidence available at the end of October 2026: two large academic studies of how ChatGPT retrieves and cites pages, the first critical survey of the field, controlled benchmarks of “GEO” content edits, a repeated-prompt experiment, a branded-citation audit, a behavioural panel with placebo controls, and two consumer surveys.
McKinsey’s June 2026 survey of 11 markets found that 53% of adults had used AI for at least one shopping task in the past three months (October 2026 analysis of about 21,000 consumers). That is self-reported use, not a share of purchases made inside AI tools. [11]
What the evidence says, and what to do about it
Within a website, what the engine’s rewritten searches retrieve explains far more of citation than FAQ blocks, schema or page speed. [1] [2] Do: map the searches engines run for your questions, then make the page that should answer them match those searches.
ChatGPT writes brands it already knows into its searches and names some brands with no supporting page. [1] Do: treat how widely and consistently a brand is described across the web as part of visibility, not only its own pages.
Lists and their order change almost every run, but the brands that appear often keep appearing. [3] Do: report “named in X of N answers” per engine, never a single rank or screenshot.
Benchmarks show many content edits do nothing or harm when others adopt them; relevance and position in context are the reproducible levers. [4] [6] Do: test changes against a baseline instead of applying a checklist.
After a recommendation, people search for the brand and visit its site; the path is mostly through search. [9] Do: track branded search and “how did you hear about us” next to AI visibility.
“AI visibility” is four different outcomes
In classic search, analysts can see a ranked results page for one query at one moment. Generative engines add hidden steps: the assistant rewrites the question into its own searches, retrieves a set of candidate pages, cites some of them, and writes a recommendation that may name brands it never cites. Jiang and colleagues traced this process in four production ChatGPT Web settings (GPT-5.2 to 5.5, Instant and Thinking). Even combined, public Google results and the OpenAI Responses API recovered only about a quarter of the exact URLs ChatGPT cited. [2]
Retrieval
Does the page enter the engine’s candidate set for the searches it actually runs?
Citation
Does the engine visibly cite the page or domain?
Recommendation
Is the brand named as an option in the answer?
Demand
Does exposure lead to searches, visits or enquiries?
The four stages are a way to organise measurement, not a proven funnel: a brand can be recommended from model memory with no citation, and a citation does not guarantee a recommendation. External tools see final answers and citations; retrieval can be observed only where an engine exposes its searches, and inferred elsewhere. [1] [2]
The strongest new evidence points upstream of page formatting
In October 2026, Mahe Chen and Nick Koudas published a study of 158,099 answers and 468,037 citations from two 90-day collections of commercial questions in contrasting markets. Its design is what makes it strong: by comparing pages within the same website, it separates the effect of a page feature from the general prominence of the site. [1]
Three results matter. First, FAQ blocks, structured data, summary boxes and page speed looked effective only when comparing different websites; within a site, page properties explained little. Pages whose titles and paragraphs matched the questions people ask were cited more. Second, when the authors replayed the questions through the engines’ APIs, every page ChatGPT cited was one its own search had retrieved, and retrieval for the sub-queries the model rewrites on every answer explained 32% of within-site variation, against 2.8% for page properties. Those rewritten queries rarely matched Google’s or Bing’s results. Third, the model’s own knowledge shapes the outcome: ChatGPT writes brands it knows into its searches, prefers websites it knows, names brands without any supporting page, and in controlled tests brand names alone reordered its rankings of products with identical facts. [1]
Jiang et al. add the production view. Instant settings mostly ran a single round of searches; GPT-5.4 Thinking often ran several. Pages retrieved more often across repeated runs were more likely to be cited. [2] The practical consequence: ranking in Google for the literal question is no guarantee of being retrieved, because the engine searches its own rewritten sub-questions.
What about the well-known “GEO” content edits?
The original GEO paper (Aggarwal et al., KDD 2024) reported visibility gains of up to about 40% from edits such as adding statistics, quotations and citations, in a benchmark where the source was already in the engine’s context. [5] It did not test whether those edits get a live page retrieved. C-SEO Bench (NeurIPS Datasets & Benchmarks 2025) tested such methods across two tasks, six domains and competing adopters: most were ineffective or harmful, improving a document’s position in the context worked better, and gains shrank as more competitors used the same methods. [6] Martinez’s critical survey of 45 studies reaches the same view: topical relevance and context position are the most reproducible levers, generic heuristics transfer poorly, and citation-oriented rewrites can even impair retrieval. [4]
Which sources AI relies on depends on the question
For an integration detail, the brand’s own documentation is the most direct evidence. For “which firm should I shortlist,” comparisons, directories and independent reviews carry more weight. A source-share percentage only means something with its denominator, engine and question type stated.
Omniscient Digital analysed 23,387 cited sources from 240 prompts that named a brand, across four industries (B2B SaaS, consumer products, professional services, finance) and six intent types. About 48% of citations were earned media, 30% other companies’ content and 23% the brand’s own content (rounded). Customer-review questions leaned on earned sources; functionality and integration questions used the brand’s own content more. [8]
Two cautions. These shares come from prompts that name the brand; unbranded “best X for Y” questions, where most new customers are won, can draw on a very different mix, and that mix varies by sector. And no study shows that appearing on a given third-party page produces a recommendation: pages differ in relevance, independence and which engines use them.
From backlinks to evidence coverage
A better question than “how many mentions do we have?” is “which buyer questions lack trustworthy evidence, and where does the evidence that engines use currently live?” For a pricing question, that is a current pricing page and corrected outdated listings. For a shortlist question, it is the directories, comparisons and reviews the engine actually cited. This question-to-evidence mapping is our working framework, built from the source patterns above; it is tested by re-measuring, not assumed.
Stale facts are a maintenance problem
Fang and colleagues attached artificial publication dates to otherwise identical passages and reranked them with seven language models. Fresher dates were consistently promoted, moving the top ten’s mean publication year forward by up to 4.78 years and single passages by up to 95 ranks; larger models reduced the effect but none removed it. [7] The lesson is not to change dates. It is to keep the facts engines rely on current, on your own pages and on the external pages that describe you.
One AI answer is not a ranking, but presence can be measured
SparkToro’s experiment (Fishkin and O’Donnell, January 2026) had 600 volunteers run 12 recommendation prompts 60–100 times each in ChatGPT, Claude and Google’s AI, for 2,961 responses. ChatGPT and Google’s AI had less than a 1-in-100 chance of giving the same list of brands twice, and about 1 in 1,000 of giving it in the same order. [3]
The same study found the opposite for presence. One cancer centre appeared in 69 of 71 ChatGPT answers while being the first mention in only 25. Across 994 responses to 142 different human-written headphone prompts, four brands appeared in 55–77% of answers. The authors conclude that visibility percentage across many prompts, run multiple times, is a reasonable metric, while ranking position is not. [3]
Put together, the evidence supports one way of measuring and rules out another. A position in one generated list is not a stable rank. The share of sampled answers that name or recommend a brand, per question group and engine, is a usable measure, provided the counts behind it are shown and each engine is reported separately.
Counts matter because small samples mislead. A 40% recommendation rate from five runs is two answers; from 100 runs it is a far steadier estimate (though runs and wordings are not fully independent). Every headline percentage should come with its numerator, denominator, engine, setting, dates and question coverage.
answers that name or recommend the brand ÷ answers sampledanswers citing a brand-owned page ÷ answers that show citationsmaterial claims about the brand that are correct ÷ claims checkedThese are the definitions Orbie uses in its reports, not industry standards. A citation of the brand’s domain is not the same as a recommendation of the brand.
Recommendations lead to demand that referral reports miss
The best evidence so far on what a recommendation does comes from Iannelli and Ai (working paper, revised August 2026). They joined opt-in web browsing data with the same users’ ChatGPT, Claude and Gemini conversations. Raw before/after comparisons overstate the effect, because many mentions are incidental references to brands the user already uses. Using matched backward placebos, a stance classifier, non-customer conditioning and a same-category control within each response, they estimate that a recommendation to a user with no recent engagement with the brand raises same-name Google searches by 4.3 percentage points, visits to the brand’s site by 2.4, and brand-specific retailer-page visits by 1.0. Incidental name-drops moved behaviour far less (+1.8 / +1.1 / +0.3). The path ran mostly through search. [9]
Two consequences follow. AI referral traffic in analytics captures only part of the effect, because most people who act on a recommendation search for the brand rather than click a link. And no single branded-search increase can be credited to one placement or page change without a comparison. McKinsey’s surveys give the context: in its five-country survey of 4,863 consumers, 28% of Gen Z and 16% of baby boomers said they used generative AI for shopping. [10] [11]
A 90-day protocol for brands and agencies
This is the protocol Orbie follows, derived from the evidence above. The numbers in it are practical defaults, not validated thresholds. Agencies can run it per client and report it in the same format.
Define the questions, measure, and record everything
- Choose 20–40 questions real buyers ask: category discovery, alternatives, comparison and verification, each in at least two natural wordings, scoped to the market the business actually serves.
- Run each question several times per engine over several days. Keep raw answers, timestamps, the searches the engine ran, cited URLs, and the answers where the brand is absent.
- Record mention, recommendation, citation, cited domain and factual errors as separate fields.
- List the facts that matter (prices, services, locations, credentials) and find where external pages contradict them.
Deliverable: a baseline with counts per question group and engine, and a register of sources and wrong facts.
Close the evidence gaps that the baseline shows
- For each question where the brand is missing, identify the likely cause: the engine’s searches don’t reach the right page, the page doesn’t answer them, independent evidence is thin, or the facts are unclear.
- Improve one page or page group at a time so it answers the searches the engine ran, in plain terms, with the entities named.
- Where recurring third-party sources are wrong or leave the brand out, ask for corrections or a fair inclusion; disclose any sponsorship and never manufacture reviews.
- Log each change with the question group it should affect and the result that would show it didn’t work.
Deliverable: a change log with dates, URLs, expected effects and untouched comparison questions.
Re-measure and separate signal from noise
- Re-run the same questions on the same engines; keep a group of untouched questions as a comparison. Compare presence and citation rates with their intervals, not rank changes.
- Check that changes hold across wordings, runs and a later date; note platform updates, launches and campaigns that could explain them.
- Set visibility changes against branded search, enquiries and what customers say about how they found the business.
- Report what didn’t work as well as what did.
Deliverable: a before/after readout with counts, intervals and the next tests.
Example: a commercial law firm
A firm is not named when people ask for an employment solicitor for small businesses in its city. First check whether engines describe it correctly when asked about it by name: if not, the problem is basic recognition, not ranking. Then look at the searches the engines ran and the pages they cited: typically legal directories and rankings, review sites and rival firms’ service pages. If the firm’s own employment page doesn’t state who it serves and where, fix that first; if the directories that keep being cited have an outdated or missing profile, fix those. Re-measure the same questions and report “named in X of N answers” per engine, before and after.
Example: an agency reporting to a client
An agency adds AI visibility to a monthly client report. Instead of a screenshot of one answer, it shows presence per question group and engine with the counts behind it, the sources cited most often for the client’s questions, the facts that are wrong in those sources, and the changes made that month. Movements are only reported as gains when they exceed run-to-run variation and hold on the untouched comparison questions.
Measure this for a brand or a client
Orbie measures how ChatGPT, Gemini and Perplexity recommend and describe a business, with the method in this review: repeated runs, results per engine, the sources behind each answer, and actions checked by a person.
How this review was built, and how far to trust it
Scope: 11 published studies (2024–2026) on retrieval, source selection, recommendation variability and downstream behaviour. Sources were found through arXiv searches, the reference list of Martinez’s 2026 survey [4] and public literature reviews. Every figure quoted is taken from the original publication, which we read; none is taken from a secondary summary.
Selection: we preferred peer-reviewed papers, preprints and studies whose method is public, and left out studies whose headline figures cannot be checked against a published method. Where authors have a commercial interest in the result, the register says so. This is a purposeful, non-systematic review, not a meta-analysis: it does not pool effects.
Controlled studies show how documents already in context are used or reranked under given settings. They may not carry over to live engines.
Large citation datasets describe their engines, questions and dates well. Scale does not remove selection bias.
Panels show behaviour after exposure; opt-in samples and confounding limit causal claims.
Claims of guaranteed ranking or fixed mention thresholds would need replicated interventions on live engines.
Evidence register
| Source | Design / scope | Used for | Main limitation |
|---|---|---|---|
| Chen & Koudas, Oct 2026 | 158,099 answers; 468,037 citations; two 90-day commercial collections + API replay | Retrieval vs page properties; model knowledge | Preprint; two markets; API settings |
| Jiang et al., Oct 2026 | Repeated production traces, four ChatGPT Web settings (GPT-5.2–5.5) | Retrieval-to-citation; proxy mismatch | Preprint; ChatGPT only |
| Fishkin & O’Donnell (SparkToro), Jan 2026 | 600 volunteers; 2,961 responses; 12 prompts; follow-up of 142 prompts / 994 responses | List and order variation; presence stability | Small prompt set; a co-author works for an AI-visibility vendor |
| Martinez, Jul 2026 | Critical survey of 45 GEO studies | Evidence boundaries; reproducible levers | Depends on the reviewed literature |
| Aggarwal et al., KDD 2024 | GEO-bench; controlled content edits | What content edits do in a fixed context | Benchmark, not live discovery; older models |
| Puerto et al., NeurIPS D&B 2025 | C-SEO Bench; 2 tasks, 6 domains, competing adopters | Limits of content edits; context position | Benchmark, not a live field trial |
| Fang et al., 2025 (rev. Sep 2026) | Artificial dates; seven reranking models | Recency bias | Artificial dates; no traffic outcome |
| Dombrowski (Omniscient Digital), Jan 2026 | 23,387 cited sources; 240 branded prompts; 4 industries, 6 intents | Source mix by question type | Brand-named prompts only; agency study |
| Iannelli & Ai, Jun 2026 (rev. Aug 2026) | Opt-in clickstream joined to ChatGPT, Claude and Gemini conversations; placebo-controlled event study | Behaviour after a recommendation | Observational; no transactions; first author works for an AI-visibility vendor |
| McKinsey, Jun 2026 | 4,863 consumers; five countries | Context: AI use for shopping by age | Self-report |
| McKinsey, 28 Oct 2026 | About 21,000 respondents; multiple markets | Context: AI use for shopping | Self-report; not comparable with [10] |
References
- Chen, M. & Koudas, N. (2026). What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO). Preprint. arxiv.org/abs/2610.15033
- Jiang, Y., Zhang, S., Xing, A., Chang, X. & Xiao, X. (2026). Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations. Preprint. arxiv.org/abs/2610.15382
- Fishkin, R. & O’Donnell, P. (28 January 2026). NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility. SparkToro. sparktoro.com
- Martinez, O. (2026). Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026). Preprint. arxiv.org/abs/2607.14035
- Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD 2024. arxiv.org/abs/2311.09735
- Puerto, H., Gubri, M., Green, T., Oh, S. J. & Yun, S. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS Datasets & Benchmarks 2025. arxiv.org/abs/2506.11097
- Fang, H., Tao, S., Chen, N., Chang, K.-X. & Sakai, T. (2025). Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking. arxiv.org/abs/2509.11353
- Dombrowski, C. (8 January 2026). How LLMs Source Brand Information: An Analysis of 23,000+ AI Citations. Omniscient Digital. beomniscient.com
- Iannelli, M. & Ai, A. (2026). From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web. Working paper, v2. arxiv.org/abs/2606.10907
- McKinsey & Company (22 June 2026). State of the Consumer 2026: When Tech Acceleration and Cost Pressures Collide. mckinsey.com
- McKinsey & Company (28 October 2026). Who Wins When AI Does the Shopping? New Evidence on Agentic Commerce. mckinsey.com