AI Mentions, Citations, and Recommendations: What Actually Matters?

The three outcomes at a glance

The three outcomes can occur independently of one another. A page from a competitor or a review site may be the cited source for an answer that recommends you. Your own page may be cited in an answer that recommends someone else. A mention buried in a caveat ("some teams also try X, though it lacks Y") counts as a mention, but it is a long way from a top recommendation.

Citations show your content was used, not that you were endorsed

A citation is evidence of retrieval. It tells you the engine pulled your page into its answer. It does not tell you that the answer represents you accurately or favorably.

The gap between citing and supporting is measurable. In a 2023 human evaluation of four generative search engines (Bing Chat, NeevaAI, Perplexity, and YouChat), only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to (Evaluating Verifiability in Generative Search Engines). Today's systems may perform differently. The underlying point still holds: an attached link does not mean the claim beside it came from that link.

The same study found something that matters for trust. Answers that readers rated as more useful tended to have less accurate citations, a correlation of r = −0.96 across systems. The authors suspect that engines which copy or closely paraphrase their sources sound fluent while citing loosely. For a brand, the implication is uncomfortable. A confident answer that recommends a competitor can look authoritative to a buyer even when its sourcing is weak. Buyers are responding to how fluent the answer is, and they rarely inspect the citations underneath it.

Citation selection also follows its own process. It is not simply a mirror of search rankings. An audit of Google AI Overviews on health, finance, and other "your money or your life" queries found that AI-generated documents were cited more often than human-authored ones, even after controlling for retrieval rank. The gap came mainly from citations of documents that weren't among the retrieved results (Auditing Citation Behavior in AI-Generated Search Summaries). Ranking first on Google does not guarantee an AI Overview citation.

Mentions show presence without position

A mention is the easiest signal to count, which makes it tempting to report as the headline metric. On its own, though, it carries no rank, no framing, and no accuracy information. "Brand X is expensive and hard to set up" and "Brand X is the best fit for small teams" each count as one mention.

Mentions are most useful as a diagnostic. If you are mentioned often but rarely recommended, the engines know your brand but don't connect it to the buyer needs you want to win. If you are recommended in some constraint-rich prompts but missing from generic category prompts, you have a positioning foothold that you can extend.

Recommendations are the outcome closest to revenue

A recommendation has two properties that a mention lacks. It has a rank: first, third, or last of five. It also has a framing, meaning the specific buyer need it answers. Together these determine whether a buyer who asks "what should I use for this?" leaves the conversation with you on their shortlist.

The most rigorous recent work treats brand recommendation as its own measurable construct. Malthouse et al. (2026) studied six LLMs across five product categories. They measured Brand Recommendation Probability (how often a brand appears in the top results) and Mean Reciprocal Rank (how high it appears on average) (Evaluating Brand Retrieval and Ranking in LLM Recommendations). Three of their findings change strategy:

• Big brands are not automatically included. Category-only queries left out many established brands, and prominence did not closely follow conventional brand popularity.
• Prominence tracks marketplace visibility. Brands that got recommended were associated with search interest and online brand conversation. The relationship is correlational, not proven causation.
• Needs-based prompts change who gets retrieved. Brands missing from ordinary lists could still surface when the prompt included distinctive cues that matched them.

The third finding matters most for smaller brands. You may never win "best CRM." You can still win "best CRM for a five-person agency that bills hourly," provided the web clearly connects your brand to that need.

Why a single prompt check gives a misleading number

AI recommendation sets change from run to run, and they change much more when the question is worded differently. This is a property of the systems themselves, not a flaw in any one tracking tool.

A 2026 study ran about 6,000 paraphrased prompts and a matching set of reruns on OpenAI and Anthropic models (Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation). Rerunning the identical prompt produced only 0.50–0.61 overlap in which brands were recommended, measured as Jaccard similarity. Cosmetic rewordings of the same buying intent dropped overlap to 0.288. Rewordings that added a constraint dropped it to 0.135. Increasing the models' reasoning effort did not close the gap. The authors argue that the wording of the prompt, more than the buyer's underlying intent, drives which brands surface. They also argue that tracking fixed prompt strings may be the wrong unit of measurement altogether.

Engines also differ from one another. A 2025 comparison of ChatGPT, Perplexity, Gemini, and Claude found that they vary in domain diversity, freshness, cross-language stability, and sensitivity to phrasing (Chen et al., Generative Engine Optimization: How to Dominate AI Search). A visibility score from one engine, or one tool's prompt set, does not transfer cleanly to another.

The practical conclusion: a screenshot of one answer is an anecdote, not a measurement.

How to measure each outcome

The ordering below follows dependency. Each step makes the next one meaningful.

1. Define your competitor set before you look at AI answers. Build the list from your own market knowledge. If you let the answers define who your competitors are, you inherit their biases, including the "big brand bias" Chen et al. describe.

2. Group prompts by buyer intent, not by keyword. For each intent you care about, write several natural phrasings. Include at least one generic category version, one constraint-rich version ("for a SaaS startup with no dev team"), and one comparison version ("X vs Y for…"). Keep branded prompts ("what is [your brand]?") in a separate bucket. Their purpose is to check how accurately you are described, not to measure discovery.

3. Run each phrasing repeatedly on each engine. The research does not establish a minimum sample size. In practice, a handful of phrasings with several reruns each gives far steadier numbers than one run per prompt. Report ranges, not single scores.

4. Record the three outcomes separately for every answer. For each answer, note whether you were cited, whether you were mentioned, whether you were recommended and at what rank, and which competitors and sources appeared instead.

5. Report recommendation probability and mean rank per intent. Treat these as the headline numbers. Use citation and mention rates to explain why the headline moved.

Illustrative example. Suppose the intent is "project management for a small creative agency." You write five phrasings, run each three times on ChatGPT, Perplexity, and Google AI Mode, and get 45 answers. You appear as a top-three recommendation in 12 of them, which is roughly a 27% recommendation rate. Nearly all 12 come from the constraint-rich phrasings, and none from the generic ones. That tells you where your positioning is working. It also tells you that a single check on "best project management tool" would have reported zero.

How to improve recommendations: work off-site first

Recommendations depend mostly on what the rest of the web says about you. Your own site plays a smaller role. Chen et al. found that AI search engines show "a systematic and overwhelming bias" toward earned media, meaning third-party and authoritative sources, over brand-owned and social content. Google's organic results draw on a more balanced mix. Malthouse et al. separately link recommendation prominence to search interest and online brand conversation. Both findings point in the same direction.

Three levers follow from this evidence:

• Get into the sources engines already cite. Look at which review sites, roundups, and publications appear in the answers for your target intents, and pursue inclusion there. Whether third-party listicles are real GEO or a new form of link building is a fair question, and it is covered in Are Third-Party Listicles Good GEO or Just AI-Era Link Building?.
• State your fit for specific needs explicitly and consistently. Because constraint-rich prompts reshuffle recommendation sets, make it clear on your site and in third-party coverage which buyer situations you are built for. Vague positioning gives engines nothing to match against a constrained prompt.
• Answer the exact buyer question directly. Research summarized by Chen et al. suggests that LLMs weigh textual relevance to the query over stylistic credibility cues (Wan et al., 2024). A page that plainly answers "is X good for agencies?" beats one that merely sounds authoritative.

One caution. In a controlled experiment with a fictitious product catalog, inserting an optimized adversarial text sequence into a product page raised its chance of becoming the top recommendation (Kumar & Lakkaraju, 2024, as cited in Chen et al.). That result describes a vulnerability, and you should not treat it as a tactic. It is also a good reason why engines, and buyers, may come to lean even harder on independent corroboration.

How to improve citations: make your content easy to draw on

Citations are the outcome that owned content affects most directly. The foundational GEO study tested content changes on a 10,000-query benchmark. Adding citations, quotations from relevant sources, and statistics raised a source's visibility in generative answers by up to 40%, and by up to 37% on Perplexity. Which strategies worked varied by domain (GEO: Generative Engine Optimization).

Read that result precisely. It measures how prominently a source features in the answer, on 2023-era engines. It does not measure whether a brand gets recommended or earns revenue, and "up to" figures are maximums, not typical gains. It is good evidence that concrete, sourced, quotable content gets drawn on more. It is not evidence that citations turn into recommendations.

A useful test for any page is to ask whether an engine could lift one sentence from it that directly answers a buyer's question and includes a specific fact. If not, rewrite until it could.

Connecting AI visibility to traffic and revenue

Do not expect citations to replace lost search traffic. In a Pew Research Center field study of Google searches, AI summaries appeared on about 18% of queries. Link clicks fell to 8% when a summary was present, compared with 15% without one. Only about 1% of clicks happened inside the AI summary itself, and about 26% of those searches ended with no click at all (Pew Research Center, 2025, as cited in Chen et al.). These figures describe Google AI Overviews, not chat assistants. They still make the core point: the answer itself is where the buyer forms an impression, and clicks are secondary.

This is why recommendation matters more than citation. If buyers rarely click through, then whether the answer names you as the right choice does most of the work.

To connect visibility to outcomes, keep each signal in its own box:

Expect measured AI referrals to undercount real AI influence. Referrals can only be recognized when the visit carries identifiable source information, so some AI-originated visits will always look like direct or unknown traffic (Tideflow AI analytics documentation). Treat referral and conversion data as a floor, and read it alongside recommendation trends.

Where Tideflow AI fits

The hard part of this framework is operational. Keeping citation, mention, rank, and competitors separate across engines and many samples is a lot of bookkeeping. So is linking those answers to what happens on your site. Doing it by hand in spreadsheets tends to collapse the distinctions this guide depends on.

Tideflow AI tracks, for each prompt across ChatGPT, Gemini, Perplexity, Google AI Mode, and Google AI Overviews, whether your brand was mentioned, whether it was cited, its rank, and which sources and competing brands appear instead. It treats crawler visits as discovery signals only, not as proof of indexing or inclusion (AI bot detection). It attributes AI referrals and conversions to the session's landing page. Visibility gaps then feed into content plans you review and approve before anything is published. These are first-party product descriptions. The platform is in early access, and like every tool in this category, it cannot guarantee recommendations. If you are comparing options, see Tideflow AI vs Profound and Tideflow AI vs Peec AI.

Frequently Asked Questions

Do ChatGPT, Perplexity, Gemini, and Google AI Overviews need different strategies?

Partly. Chen et al. found that these engines differ in how diverse and how fresh their sources are, and in how sensitive they are to phrasing. So measure each engine separately rather than averaging them into one score. The shared foundation is the same across all of them: earned third-party coverage and clearly stated fit for specific buyer needs.

Should we publish AI-generated content because AI Overviews cite it more?

No. The AI Overviews citation audit covered only YMYL queries, and its authors do not recommend producing AI-generated content. The finding shows that citation selection departs from search rank. It does not show that machine-written pages earn recommendations or trust.

Is mention volume a reasonable KPI to report to leadership?

Only as a supporting metric. Mentions carry no rank or framing, so a rising mention count can hide a falling recommendation rate. Report recommendation probability and mean rank per buyer intent as the headline numbers, and use mentions and citations to explain the changes.

Your next step

This week, pick your three most important buyer intents. Write four or five natural phrasings for each, including at least one with a specific constraint, and run them several times on two or three engines. Record citation, mention, and recommendation with rank separately. The pattern of where you are recommended, and where you are only mentioned or absent, tells you whether your next move is third-party coverage, sharper positioning, or better-sourced content. For a hands-on walkthrough of improving your presence in one engine, continue with How to Get Mentioned by ChatGPT: A Tested, Step-by-Step Playbook.

Sources

• arXiv, Evaluating Verifiability in Generative Search Engines; Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation: Reproducibility Below the Rerun-Stability Baseline (May 2026); GEO: Generative Engine Optimization
• Proceedings of Machine Learning Research, Auditing Citation Behavior in AI-Generated Search Summaries: A Framework and a Case Study of Google AI Overviews (June 2026)
• Malthouse et al., Evaluating Brand Retrieval and Ranking in Large Language Model Recommendations (September 2026)
• Chen et al., Generative Engine Optimization: How to Dominate AI Search
• Tideflow AI, Add analytics and track views and conversions; AI Visibility & Content Platform; Detect AI Bots