SourceSight
Insights

AI search research

Topic
AllHow engines workSourcesIndustriesMethods
Industry
AllLocal services2B2B & export3SaaS & tools1E-commerce8Media1Regulated9Travel & hospitality4Education3Auto & real estate1Pet care1Food & beverage1Logistics2Consumer & household5Support & after-sales1Professional services1Hiring & HR1Home services1Insurance1
IndustriesE-commerceConsumer & household2026-10-07

Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages

Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.

Setting: recommending general merchandise to grocery-focused shoppers (pairing milk with a milk frother) is a long under-explored cross-category problem requiring product associations and real-time cart context.

Key numbers · 4References · 1Read →
IndustriesConsumer & household2026-10-07

Handing appliance scheduling to AI: it captures 96.7–98% of the savings — and still needs a validator

Five LLMs acting as home energy managers: cheap is not the same as safe or feasible, models both invent schedules and explain a constraint correctly while committing an invalid one.

System: a tool-calling ReAct agent converts natural-language requests into half-hourly schedules for multiple household appliances, using dynamic tariffs, weather forecasts, PV generation estimates and household demand data.

Key numbers · 5References · 1Read →
Methods2026-10-07

Same engine, different surface, different answer: test app, web and API separately

On the same platform and the same questions, app and web cite systematically different source sets — and the API shows something else again.

Chinese generative search: four mainstream platforms across web and app — eight interfaces, 614 queries, three replications per combination, cleaned from 214,119 raw records into 160,860 citation-level records; one conclusion is that app and web source sets differ systematically for the same platform.

Key numbers · 4References · 3Read →
IndustriesSupport & after-salesE-commerce2026-10-07

When service teams get AI: faster and better rated — but retrials unchanged, and top agents get worse

A large-scale field experiment in Alibaba's after-sales chat support: the assistant only drafts opening diagnoses and proposals, agents can adopt, modify or ignore — and who is using it decides the outcome.

Design: a large-scale field experiment with Alibaba randomised whether human chat agents could access a generative AI assistant that drafts issue diagnoses and solution proposals in the opening stage only; agents could adopt, modify or disregard them, so the study estimates effects of both access and usage.

Key numbers · 4References · 1Read →
Methods2026-10-07

How to design a retest you can trust: five moves so noise does not pass as progress

AI visibility is a distribution, not a point: fix the prompts, repeat the runs, add controls, paraphrase, validate by hand — or you cannot attribute the change.

Why repeats are mandatory: one GEO measurement study concludes answers vary across runs, prompts and time, making one-off observations unreliable and visibility a distribution; a brand-recommendation evaluation study likewise requires repeated sampling because the same query yields different brands and rankings.

Key numbers · 4References · 3Read →
Methods2026-10-07

What GEO methods actually hold up: 45 studies and a cold shower

The widely cited “40% visibility lift” comes with a condition: the source was already in a fixed context. It proves neither discoverability nor durable traffic.

Scope: 45 studies published between November 2023 and July 2026 (plus one earlier preprint accepted at EMNLP after the window opened), together with related RAG and evaluation work.

Key numbers · 5References · 1Read →
IndustriesEducation2026-10-07

AI career advice is generic: local context covered only 35.4% of the time across ten countries

The “described correctly” problem applies to education too: models know the tech stack but frequently skip local ecosystems, language requirements and national policy.

Design: standardised prompts tested across six models (ChatGPT-4, DeepSeek, Gemini, Claude 3.5, Llama 3, Mistral) describing entry-level computing careers in ten African countries — 60 responses analysed against the Computing Curricula 2020 framework.

Key numbers · 4References · 1Read →
IndustriesTravel & hospitality2026-10-07

Which hotel does AI recommend? Rating +31.6pp, high price −30.0, list position worth $12 a night

A randomised audit quantified the signal weights behind AI hotel picks: rating and price dominate, eco-certification is overweighted and management responses are ignored.

Design: a pre-specified choice-based conjoint audit across personas, prompt templates and twelve open-weight and proprietary models, asking assistants to choose among five hotels with guest rating, review volume and recency, management response, chain affiliation, price, eco-certification and list position independently randomised.

Key numbers · 4References · 1Read →
IndustriesLogistics2026-10-07

Multi-warehouse allocation: letting an LLM pick the formulation lifted hit rate from 21% to 50%

No single mathematical formulation wins every allocation instance, so choosing the right one per case beats reusing the best fixed model.

Problem: multi-warehouse inventory allocation is usually formulated as mixed-integer programming, but no single formulation consistently matches every regime — demand concentration, inventory imbalance, replenishment scale, service constraints and forecast volatility all shift what is optimal.

Key numbers · 4References · 1Read →
Sources2026-10-07

What Chinese generative search actually cites: 8.3% brand selection, and app vs web cite different sources

Across 160,860 citation-level records, the same platform surfaced systematically different source sets on its app and web interfaces.

Scale: four mainstream Chinese platforms across web and app — eight interfaces, 614 queries, three replications per query-platform-interface combination; 214,119 raw records cleaned into 160,860 citation-level records.

Key numbers · 5References · 2Read →
IndustriesB2B & export2026-10-07

Unstructured supplier data means you do not exist to AI

AI finds four times as many compatible suppliers as keyword search across 10,000 records — and still misses you if the capacity you offer is not written down.

The starting point: when supply chains are disrupted, capacity outside the network is hard to use — incumbents' models are private, provider profiles are unstructured, and offers stay hidden until costly engagement.

Key numbers · 5References · 1Read →
IndustriesRegulated2026-10-07

Same dental case, four different answers: DeepSeek came out on top

In specialist healthcare, finding out which engine describes your field accurately matters more than whether “AI mentioned us”.

Models evaluated: four frontier systems (GPT-4o, Gemini 2.0 Flash, Copilot and DeepSeek V3), tasked with open-ended clinical analysis of longitudinal dental cases.

Key numbers · 4References · 1Read →
IndustriesRegulated2026-10-07

High-stakes AI advice shifts with conversation context — not with who you are

For legal, medical and financial questions, a single conversation history is enough to give different users different advice. The strongest predictor is the topic.

Background: in high-stakes settings such as legal, medical and financial advice, even a single conversation history is enough to drive outcome differences between users, and prior work shows socio-demographic groups receive advice of differing quality.

Key numbers · 4References · 1Read →
How engines work2026-10-07

What 2 million AI citations show: prompt-content match matters most, checklists matter less

Once you control for the domain, the usual AEO checklist (FAQ blocks, structured data, Core Web Vitals) reverses or collapses to zero — domain authority and prompt alignment dominate.

Sample: roughly 2 million citations from four commercial engines (including ChatGPT, Claude and Gemini) over six months, joined to 10,000 crawled pages across 19 B2B SaaS workspaces, with 60+ page features tested.

Key numbers · 5References · 1Read →
IndustriesB2B & export2026-10-07

B2B buyers start with AI, then verify with sales: 13 internal stakeholders and 9 external influencers per purchase

Answer engines have become the starting point for B2B buying, but buyers do not fully trust them — they go find a human or a third party to validate.

Forrester, The State of Business Buying, 2026: a typical B2B purchase now involves 13 internal stakeholders and nine external influencers, and the numbers grow with complexity and strategic importance.

Key numbers · 5References · 2Read →
How engines work2026-10-07

Whose side is the AI shopping assistant on? Change its role and the penalty for sponsored results disappears

The same sponsored listings get treated very differently once the prompt names the platform as the agent's principal — disclosure rules built for humans do not carry over.

Controlled choice experiments varied only the assigned role in the system prompt (naming either a traveller or a booking platform as the agent's principal) and compared how the model handled sponsored listings and reasoned about them.

Key numbers · 4References · 1Read →
How engines work2026-10-07

AI search and Google are not the same: identical content, completely different sources

AI search systematically favours third-party authoritative sources while Google keeps a more balanced mix — lifting an SEO playbook across will not work.

Controlled experiments across verticals, languages and query paraphrases compared AI search (ChatGPT, Perplexity, Gemini, Claude) with Google: AI search shows a systematic, overwhelming bias toward earned media over brand-owned and social content.

Key numbers · 4References · 1Read →
IndustriesEducation2026-10-06

Before handing study-abroad advice to AI: separate “wrong” from “said too much”

For high-stakes admissions, visa, scholarship and eligibility questions, models tend to under-cover or over-reach when domains intersect — so evaluation has to score accuracy and hallucination together.

Use case: admissions, visa, scholarship and eligibility questions — high-stakes study-abroad topics increasingly answered directly by LLMs, with unclear reliability beforehand.

Key numbers · 4References · 1Read →
Methods2026-10-06

How to evaluate a conversational AI system: 23 metrics plus agreement with human raters

Single-turn, single-task accuracy cannot capture real use; multi-turn dialogue needs five dimensions, 23 metrics — and proof that machine scoring matches human scoring.

Framework: derived using the Goal-Question-Metric method across five categories — task performance, factual accuracy, interaction quality, control capability and system efficiency.

Key numbers · 4References · 1Read →
Methods2026-10-06

How to build a per-market source list: see who a market reads before deciding what to fix

The top cited source differs by market — YouTube in Poland, a local business daily in Lithuania — and Chinese platforms differ between app and web.

Cross-market evidence: 128 brands, 12 home markets, 13 languages and 167,551 URL-grounded citations — Wikipedia was the most-cited domain in 11 of 12 languages, the exception being Lithuanian, where the business daily vz.lt led at 4.38%.

Key numbers · 4References · 2Read →
IndustriesE-commerce2026-10-06

When AI agents spend money: 8 of 20 fraud classes are no better than guessing

The biggest agentic payment risk is not impersonation: the counterparty, settlement address and delivered service are all genuine — it just charges too much, and no identity check can see it.

Setting: AI agents already hold spend authority and settle payments without per-action human confirmation; the resulting loss is often not a security failure but a counterparty that is exactly who it claims to be, delivers the service, and overcharges.

Key numbers · 5References · 1Read →
IndustriesProfessional servicesRegulated2026-10-06

Once AI makes expertise cheap to fake, what still signals competence? Enforceable promises, not badges

When AI produces convincing reports and proposals at near-zero cost, signals built on costly effort stop working — the paper's substitute is a verifiable promise backed by damages.

Framework: generative AI is modelled as compressing the discernible headroom between what machines produce cheaply and what buyers can distinguish; below a critical headroom no separating equilibrium in production-side signals exists.

Key numbers · 5References · 1Read →
SourcesConsumer & householdE-commerce2026-10-06

Incumbent advantage: with identical specs, a well-known brand is recommended 100% of the time

Three experiments on skincare: identical specifications produce near-total dominance for known brands — and less than a +0.1-star edge for a competitor breaks it.

Setup: skincare, a category where consumers cannot judge quality before buying and must rely on reputation, tested across three commercial models (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash), with a robustness check on search goods.

Key numbers · 4References · 1Read →
Sources2026-10-06

Does AI judge content or clout? A causal test of authority bias

Hold the title and abstract constant, change only author prestige, venue and citation counts, and eight models change their recommendation — while claiming they do not.

Design: authority metadata — author prestige, venue and citation counts — varied across three counterfactual conditions (original, flipped, boosted) with titles and abstracts held constant.

Key numbers · 4References · 1Read →
IndustriesRegulated2026-10-06

Healthcare AI's old problem: it knows the guideline but cannot cite it, let alone follow it

Across 32,155 clinical recommendations, models detected 71.1–89.6% but correctly referenced the source title only 3.6–29.7% of the time, and adherence ranged from 21.8% to 63.2%.

Scale: 3,418 clinical practice guideline documents from nine countries/regions and two international organisations over the past decade, spanning 24 specialties, from which 32,155 clinical recommendations were extracted with publication body, date, country, specialty, recommendation strength and evidence level.

Key numbers · 4References · 1Read →
IndustriesTravel & hospitality2026-10-06

AI now reads the reviews for you: summaries plus targeted questions replace scrolling

The decisive information in accommodation decisions lives in unstructured reviews, and summarising them with query support noticeably shortens the search.

Context: booking platforms such as Booking.com and Airbnb expose enormous volumes of data; structured filters cover stars, amenities and price, but the most valuable signals sit in unstructured review text.

Key numbers · 4References · 1Read →
IndustriesLogistics2026-10-06

Real freight-network risk costs more than history suggests: +25%, not +3%

Assessing resilience with historical disruption scenarios understates systemic risk by an order of magnitude; generated scenarios push expected annual cost to $5.11M.

Problem: intermodal freight networks face correlated, multi-mode disruptions, yet resilience assessments often rely on historical or uncorrelated scenarios that understate systemic risk.

Key numbers · 4References · 1Read →
IndustriesFood & beverage2026-10-06

Food delivery recommendation is now two-stage: guess the craving, then rank the vendors

For repeat orders, ranking is no longer pure CTR prediction: cuisine prediction comes first, vendor ranking second, and the LLM works in the re-ranking step.

Framework: MARS splits repeat-order recommendation into two stages — cuisine prediction, then vendor ranking — combining LightGCN global preference signals, Swing-based local peer evidence, geospatial filtering and prompt-driven LLM reasoning over behavioural, temporal and geographic context.

Key numbers · 3References · 1Read →
IndustriesB2B & export2026-10-06

Supply-chain risk analysis: from days to 3.83 minutes

Signals across multi-tier suppliers used to take analysts days to assemble; an agent pipeline now returns an end-to-end analysis in 3.83 minutes at $0.08 per disruption.

Background: modern supply chains face geopolitical events, demand shocks, trade restrictions and natural disasters, and many disruptions originate deep in the network while most companies still cannot see past Tier-1 suppliers.

Key numbers · 4References · 1Read →
IndustriesEducation2026-10-06

Students are asking AI which major to choose: fix the bias first, then talk accuracy

The front door to admissions and academic advice is becoming a chat window, and AI answers can carry stereotypes — a risk and an opportunity for institutions.

Context: choosing a major is hard for incoming students, and traditional advising suffers from long waits, intimidating environments and limited personalisation — which is where chatbots step in.

Key numbers · 4References · 1Read →
IndustriesRegulated2026-10-06

In law, the gap is not legal knowledge but wrong jurisdiction and wrong citations

The same facts can lead to different conclusions in China, California and Germany; models often answer correctly yet fail to cite the right jurisdiction.

Benchmark scale: three jurisdictions (China, California, Germany), 55 legal issues across contract, consumer, criminal, family and labour law, 6,149 instances grouped into 385 fact patterns, all reviewed by legal professionals.

Key numbers · 3References · 1Read →
Sources2026-10-06

Ask the same question in another language and the answer changes: what cross-lingual consistency research shows

Ask in Chinese, Traditional Chinese or English and AI often answers differently — a documented model-level property, not a one-off glitch.

Multiple studies converge on the same phenomenon: multilingual models give inconsistent answers to semantically equivalent questions across languages, which is why cross-lingual consistency (CLC) has become its own research area.

Key numbers · 4References · 2Read →
How engines work2026-10-06

Is your domain in the retrieval path? That's the difference between 2.8% and 91.4% mention rate

When neither the brand nor its own domain appears in the retrieval path, mention rates sit at 2.8% (GPT) and 3.8% (Gemini). With an own-domain citation, the same prompt reaches 49–58%.

Sample: 34,960 unbranded prompt-engine observations from 75 projects, covering 2,854 distinct prompts, with repeated GPT and Gemini runs from June to September 2026.

Key numbers · 4References · 1Read →
Methods2026-10-05

How to retest off-site coverage: register, re-run the same prompts, compare the source pool

Off-site work is not finished when you publish: re-run the same prompts, check whether the cited source pool changed, and whether the new sources brought discoverability.

Registration has quantified value: across 200 synthetic disruption episodes, registering offers in the same index lifted captured value from 46.8% to 98.6% — unregistered capacity caused requests to fail outright.

Key numbers · 4References · 3Read →
IndustriesConsumer & householdE-commerce2026-10-05

Grocery recommendations move from items to categories — and cart-adds rise 7.5%

Online grocery has cyclical repurchase patterns that item-level models struggle to scale on; category-level recommendations with constrained decoding lifted production cart-adds by 7.5%.

Problem: online grocery recommendations must capture cyclical buying behaviour and diverse intents, while traditional item-level methods face scalability and accuracy limits — category-level recommendation is more structured and practical.

Key numbers · 4References · 1Read →
Methods2026-10-05

Small businesses can run their own Q&A assistant too: low-cost multi-tenancy and injection defence

For small teams the hard part is not model choice but satisfying isolation, cost and security at the same time.

Setting: an open-source multi-tenant platform that lets small businesses deploy customised LLM support chatbots through a no-code workflow, where the obstacles are infrastructure cost, engineering complexity and RAG-specific security risk.

Key numbers · 5References · 1Read →
Sources2026-10-05

Web search and AI answers differ less in format than in who they read

For the same queries, Google results and generative AI answers diverge systematically in source domains, source types, query intent and freshness.

Scale: a large-scale empirical study comparing results returned by Google Search against answers from leading generative AI services.

Key numbers · 4References · 1Read →
Methods2026-10-05

Stop collapsing visibility into one score: separate discoverability, citation, absorption and revenue

A single “AI visibility score” hides four different links in the chain: being retrieved, being cited, being absorbed into an answer, and producing business.

The survey decomposes GEO into a multistage pipeline: search activation → crawling and indexing → retrieval → reranking and context allocation → citation → prominence → factual absorption → fidelity → user behaviour.

Key numbers · 4References · 1Read →
Sources2026-10-05

National bias across languages: ask in another language, get another answer

In university applications, travel and relocation tasks, model recommendations shift with the language — local language bias is prevalent.

Scope: three personalised-advice scenarios — university applications, travel and relocation — tested across multiple languages for decision-making tasks.

Key numbers · 4References · 1Read →
IndustriesTravel & hospitality2026-10-05

Turning reviews into an intent graph: how platforms infer what guests really want

Machines do not just read sentiment in reviews, they reconstruct unstated needs — and that drives which guests each property gets matched to.

Method: build a user-specific intent knowledge graph from each review, align it with a global hotel knowledge graph through structure-aware semantic smoothing, then let an LLM jointly predict aspect ratings and generate reverse user-intent statements.

Key numbers · 4References · 1Read →
IndustriesConsumer & household2026-10-05

Can AI predict household spending? A scorecard across 6 surveys and 32 tasks

Most models beat a no-change baseline, but household predictions still carry systematic over- and under-prediction — and gradient-boosted trees still lead on most tasks.

Benchmark scale: six US household surveys unified into 32 prediction tasks spanning numeric, categorical and probabilistic outcomes across consumption, income, labour, expectations and housing.

Key numbers · 5References · 1Read →
IndustriesPet care2026-10-05

Pet owners ask AI first: what a veterinary knowledge benchmark shows

Across 18,827 expert-answered question pairs, how reliable a model is on pet clinical questions depends on the material it reads.

Benchmark scale: 10,076 text-only plus 8,751 multimodal question-answer pairs drawn from real questions about dogs and cats, paired with answers from expert veterinarians.

Key numbers · 4References · 1Read →
IndustriesRegulated2026-10-05

In finance, an LLM needs structured signals to be accurate

Judging credit risk with a model alone ranks worse than a classic model — feed the classifier's key features and probability into the prompt and the metrics catch up.

Task: credit-card default prediction on the Default of Credit Card Clients dataset, comparing a fitted classifier, LLM prompting, and combinations of the two.

Key numbers · 4References · 1Read →
IndustriesSaaS & tools2026-10-05

What AI recommends when you pick software: popularity, not fit

Hand selection to an AI and you often get a popularity chart rather than a fit for your context — dependencies get ignored and evidence stays opaque.

The study frames software package selection as multi-criteria decision-making (MCDM), with automated pipelines continuously collecting metadata, usage trends, vulnerability data and developer sentiment from GitHub, PyPI and Stack Overflow.

Key numbers · 4References · 1Read →
MethodsE-commerce2026-10-05

Can AI actually read your product pages? A 3,000-page extraction experiment

What AI needs from a product page is not pretty layout but extractable structured facts: specs, ingredients, price, certification.

Scope: 3,000 food product pages from three online retailers, with the goal of extracting key attributes such as ingredient lists and nutrition tables.

Key numbers · 4References · 1Read →
IndustriesMedia2026-10-05

AI is the new front door to news: weekly chatbot news use rose from 7% to 10%

Social media and video overtook news websites as the main source of online news for the first time — with AI chatbots the fastest-growing new entry point.

Reuters Institute Digital News Report 2026 (48 markets): weekly use of AI chatbots for news rose from 7% to 10%, reaching 16% among under-35s.

Key numbers · 5References · 1Read →
Sources2026-10-05

Ask in another language and the citations get thinner: a 15,942-citation multilingual audit

Non-English queries surface fewer citations and are routed to language-appropriate sources far less often — translation alone is not a multilingual strategy.

Audit: 20 English mental-health questions under two prompt conditions, a subset translated into six further languages — 1,140 responses, 15,942 citations and 1,713 unique domains recorded.

Key numbers · 4References · 1Read →
IndustriesTravel & hospitality2026-10-04

Travel and hospitality: AI itineraries collapse under delays and crowds

Existing benchmarks assume a deterministic world. Add real delays and crowd fluctuations and the gap between model plans and human plans becomes obvious.

Data scale: real travel data spanning 504 cities in India, including attractions, restaurants, accommodation and multi-modal transport networks.

Key numbers · 3References · 1Read →
Methods2026-10-04

How to test three languages: run Chinese, Traditional Chinese and English separately

Cross-lingual consistency research shows answers change with language, so multilingual markets must be asked and measured language by language.

Finding one: multilingual models frequently answer semantically equivalent questions inconsistently, which is why cross-lingual consistency (CLC) is its own field; direct distribution alignment in post-training improved consistency across three model families and three benchmarks.

Key numbers · 4References · 3Read →
How engines work2026-10-04

When an AI agent chooses for you: price, ratings and nudges move its decisions too

Agents have no human cognitive limits, yet they are still biased choosers — change how a price is framed, where a rating comes from or how a nudge is worded, and the choice changes.

Framework: ABxLab probes agentic choice in a realistic web shopping environment by systematically varying option attributes (price, ratings) and persuasive cues (psychological nudges).

Key numbers · 4References · 1Read →
Methods2026-10-04

Why one measurement is not a measurement: AI visibility is a distribution

Ask the same question twice and the brands and rankings change — a single observation cannot tell you where you stand.

A study on measuring GEO concludes answers vary across runs, prompts and time, making one-off observations unreliable: brand visibility should be characterised as a distribution, not a single point.

Key numbers · 4References · 2Read →
IndustriesAuto & real estate2026-10-03

Long-cycle industries like auto and real estate: AI is moving from answering questions to deciding who gets looked at first

In long, multi-stage funnels, lead ranking is worth more than an impression — one NEV brand lifted top-lead precision by 39.7%.

Industry context: lead conversion in high-ticket sectors such as automotive and real estate differs fundamentally from e-commerce recommendation — prolonged decision cycles and multi-stage funnels.

Key numbers · 4References · 1Read →
IndustriesRegulated2026-10-02

How regulated industries get cited: compliance signals are authority multipliers — if machines can read them

In heavily regulated sectors like UK iGaming, AI visibility is no longer decided by keyword density but by algorithmic trust — and compliance facts have to be machine-readable.

An empirical analysis of UK iGaming entities finds that compliance signals (such as UK Gambling Commission standards) act as authority multipliers for models like ChatGPT, Perplexity and Gemini once they are properly structured.

Key numbers · 3References · 1Read →
How engines work2026-08-10

61.7% of AI citations never name the brand

AI cited your page, but customers never saw your name. That is the norm, not the exception.

Across 3,981 domains, 115 prompts, 14 countries and 4 engines, an industry analysis found 61.7% of citations never name the brand.

Key numbers · 4References · 1Read →
How engines work2026-08-10

Relevance and position decide who gets cited first

252,000 controlled trials: appear early and answer the question, get cited.

A controlled RAG testbed ran 252,000 paired trials across six LLMs, changing one content factor at a time.

Key numbers · 3References · 1Read →