Ask ChatGPT, Google AI Mode, Gemini, Perplexity or another AI service to recommend a business in your market. Does your company appear? Which competitors appear? How is each business described? What evidence supports the answer? AI can increasingly help somebody understand a problem, find providers, compare options and decide what matters before they reach a supplier website.
Research boundary
This article separates direct evidence from strategic interpretation. WLW’s 191 homepage benchmark describes homepage SEO, performance, security and UX. It does not measure ChatGPT rankings or prove AI recommendation performance. Google, OpenAI and Anthropic research are used as external evidence. AI recommendation testing is treated as directional research rather than a fixed ranking.
This article separates direct evidence from strategic interpretation. WLW’s 191 homepage benchmark describes homepage SEO, performance, security and UX. It does not measure ChatGPT rankings or prove AI recommendation performance. Google, OpenAI and Anthropic research are used as external evidence. AI recommendation testing is treated as directional research rather than a fixed ranking.
191matched homepage audits providing the WLW online presence evidence base
4customer intent lenses: understand, find, solve and choose
5proof types: owned, customer, independent, expert and commercial
4confidence levels: known, supported, assumed and unknown
The headline findings
The useful conclusion is not that every business suddenly needs a separate “AI SEO” programme. The stronger conclusion is that AI adds another discovery and comparison layer to an online presence which was already being shaped by search, content, proof, reviews, brand, websites, customer journeys and external authority.
Search remains a measurable foundation
Google says its generative AI search experiences remain rooted in core Search ranking and quality systems. Search demand, crawlability, useful pages and clear information architecture still matter.
AI can influence the shortlist before the visit
A customer can ask an AI system to explain a market, identify providers, compare alternatives or recommend what to consider before reaching any supplier website.
Real expertise is not automatically visible expertise
Knowledge can remain trapped inside people, projects, products, research, customer service and sales conversations without becoming useful public evidence.
Proof matters more than publishing volume
Case evidence, customer outcomes, research, expert knowledge and independent validation create more substance than a large number of interchangeable articles.
Different AI services need separate research
Google, ChatGPT Search, Perplexity and other services do not behave identically. One answer from one platform should not be treated as universal AI market share.
Commercial value outranks AI vanity metrics
Qualified enquiries, pipeline, sales, customer behaviour and first party search evidence should carry more weight than a directional AI visibility score on its own.
Use four confidence levels before making an AI recommendation
The v2.1 WLW framework prevents observations from turning into claims without enough evidence.
KnownDirect evidence establishes the conclusion
SupportedSeveral credible signals point in the same direction
AssumedA reasonable strategic proposition which still needs testing
UnknownThe evidence is not strong enough to conclude
Start with the commercial question, not the AI tactic
A useful AI visibility strategy starts with the same question as a useful search, content or website strategy: what commercial or organisational outcome are we trying to improve?
Traffic is not automatically the objective. Being mentioned by an AI system is not automatically the objective either. A manufacturer may need more technically qualified enquiries. A retailer may need customers to choose the correct product more confidently. A professional services business may need to become visible for a problem before the prospect has chosen a provider.
DemandWhat problem, requirement, opportunity or event creates the need?
LanguageWhich keywords, queries and customer words express that demand?
DecisionWhat determines suitability, confidence and choice?
EvidenceWhich proof and authority support the claims being made?
OutcomeWhat action creates commercial or organisational value?
Completeness of research does not require completeness of recommendation. A strong process can research the whole opportunity then recommend only the work with a clear role in changing the outcome.
Understand, find, solve and choose
Keyword and query research remain core inputs because they reveal market language, demand, page subjects and search intent. The v2.1 framework adds four practical intent lenses to explain why that demand exists. They do not replace keyword research.
Explore the four customer intent lenses
Use the buttons to see how the same market can produce different questions and different content requirements.
Example AI question: What should a UK manufacturer consider before choosing an industrial marketing partner? The useful answer needs terminology, options, risks, decision factors and proof.
Example AI question: Which UK agencies specialise in marketing for manufacturers and engineering businesses? The useful evidence is category relevance, audience, location, services and applications.
Example AI question: Who can help a technical B2B company generate more qualified enquiries from a long sales cycle? The answer needs capability, process, constraints, experience and outcomes.
Example AI question: Compare these agencies for a manufacturer that needs research, website improvement, content and lead generation. The answer needs differences, proof, service fit, risk and reasons to choose.
Do not substitute situations for demand evidence.
Customer situations help explain the meaning behind a keyword or query. They should sit alongside search volume, query data, customer questions, analytics, sales evidence, reviews, competitor research and market evidence rather than replace them.
Customer situations help explain the meaning behind a keyword or query. They should sit alongside search volume, query data, customer questions, analytics, sales evidence, reviews, competitor research and market evidence rather than replace them.
What AI visibility actually means
AI visibility is the broad question of whether a business, brand, product, page or source appears or contributes when people use generative AI to research or evaluate information. AEO and GEO are common labels, but the commercial work underneath them still needs to be defined properly.
Drag or swipe the table sideways to compare the terms.
| Term | Practical meaning | Commercial use | Check and balance |
|---|---|---|---|
| AI visibility | Whether the organisation or its evidence appears in generative AI discovery. | Broad discovery and representation measure. | Do not reduce it to one third party score. |
| AEO | Answer Engine Optimisation. | Useful shorthand for answer based search visibility. | Google says generative AI Search remains rooted in SEO fundamentals. |
| GEO | Generative Engine Optimisation. | Useful shorthand for generative AI visibility. | It does not replace demand, proof, authority or conversion. |
| AI recommendation visibility | Whether a company appears in recommendations, comparisons or shortlists. | Commercially important before the buyer knows which provider to choose. | Prompt testing is directional rather than a fixed ranking. |
| AI citation visibility | Whether a page or source is surfaced as supporting evidence. | Shows what material systems are using to support an answer. | A citation is not proof of revenue or preference. |
What 191 homepage audits show
The WLW UK Online Presence Benchmark gives us a useful foundation because it shows how uneven the online presence can be even before AI recommendation visibility is considered.
68.4average SEO score across the 191 matched homepage audits
53.1average performance score
50.8average security score
75.4average UX score
Across the 14 industry averages, performance or security is the weakest area every time. UX is the strongest average area in 10 of 14 industries. Businesses are rarely weak everywhere. The useful question is where their strengths are being limited.
The gaps can be more revealing than the overall score
Different parts of the same online presence can be separated by more than 30, 40 or even 50 points.
42.6point gap between manufacturing SEO 72.6 and performance 30.0
51.0point gap between energy performance 75.0 and security 24.0
31.7point gap between travel UX 80.7 and security 49.0
34.5point gap between professional services SEO 72.0 and performance 37.5
Benchmark boundary
The homepage scores do not tell us how a company performs in ChatGPT Search, Google AI Mode, Perplexity or another AI service. They show why AI visibility should be researched inside the wider online presence rather than inferred from one technical score.
The homepage scores do not tell us how a company performs in ChatGPT Search, Google AI Mode, Perplexity or another AI service. They show why AI visibility should be researched inside the wider online presence rather than inferred from one technical score.
The digital evidence footprint
An organisation knows far more about itself than the internet does. Inside a business may sit years of expertise, customer questions, successful projects, product knowledge, research, data and practical lessons. None of it automatically becomes useful public evidence.
Identity
Who is the organisation, where does it operate and how should it be distinguished from similarly named businesses?
Offer
What products, services, applications and outcomes does it actually provide?
Audience
Who is the offer suitable for and which sectors, situations or requirements does it serve?
Decision factors
What determines suitability: price, capability, availability, compliance, experience, support, risk or another factor?
Evidence
What results, research, case studies, reviews, expertise and independent sources support the claims?
Action
What should the customer do next and can the website turn the resulting attention into useful value?
DiscoverableCan relevant pages and evidence be found?
UnderstandableCan the offer, audience and relevance be understood?
VerifiableCan important claims be supported?
RecommendableIs there enough relevant evidence to justify inclusion?
ConvertibleCan attention become a useful next action?
What AI may pick up is not hidden AI text. It is explicit relationships. Who the business helps. What it solves. When it is relevant. What determines suitability. What proof exists. Which experts stand behind the work. What a customer should do next.
Five kinds of proof
Authority becomes more useful when it is broken down into evidence a customer can actually evaluate.
Owned proof: case studies, results, research, technical data, processes, guarantees and accreditations
Customer proof: reviews, testimonials, customer stories, outcomes and references
Independent proof: press, industry publications, associations, directories, partners and independent research
Expert proof: named specialists, engineers, clinicians, consultants, founders, product experts and credentials
Commercial proof: revenue, cost reduction, conversion improvement, time saved, efficiency and measurable outcomes
The useful question is not “how many mentions do we have?” It is “which important claims can be supported, what proof already exists and what proof still needs to be created or earned?”
Google, ChatGPT and other AI services should not be treated as one system
Current platform evidence supports different conclusions for different services.
Google generative search
Search remains the foundation
Google says AI Overviews and AI Mode are rooted in core Search ranking and quality systems.
- Existing SEO best practice remains relevant
- Useful, unique and non commodity content matters
- Clear textual information and internal links matter
- No special AI schema is required
- Artificial content chunking is not required
- Google says llms.txt does not affect Search visibility
ChatGPT Search
Public sites can be surfaced and cited
OpenAI says any public website can appear in ChatGPT Search and OAI SearchBot access helps content become discoverable, surfaced and clearly cited.
- Placement is not guaranteed
- Search results can contain source citations
- Referral traffic can be measured in analytics
- Search crawler access is separate from model training controls
- Current web retrieval should not be confused with permanent model memory
Other AI services
Research them separately where they matter
Perplexity and other services can be commercially relevant, but their behaviour should not be assumed from Google or ChatGPT.
- Use realistic customer situations
- Record sources and descriptions
- Track competitor appearance
- Repeat tests before drawing conclusions
- Treat the results as directional evidence
Search crawling is not model training
This distinction matters because a great deal of AI marketing blends two different processes into one story.
Search discovery
A crawler can access public pages so an AI search product can discover, retrieve, summarise or cite current information.
NowTiming
WebSource
QueryTrigger
CiteUse
TrackMeasure
Model training
Training concerns whether content may be used when future models are developed. Businesses cannot infer future inclusion simply because a page is public or crawlable for search.
FutureTiming
DataSource
TrainProcess
LearnUse
UnknownInclusion
Check and balance
Allowing OAI SearchBot so content can be eligible for ChatGPT Search does not establish that the same content will be used for future model training. OpenAI publishes separate controls for search discovery and potential model training.
Allowing OAI SearchBot so content can be eligible for ChatGPT Search does not establish that the same content will be used for future model training. OpenAI publishes separate controls for search discovery and potential model training.
What the 250 document study really says
The number 250 comes from genuine research by Anthropic, the UK AI Security Institute and the Alan Turing Institute. The study tested data poisoning: whether deliberately malicious documents could create a specific backdoor behaviour in experimental language models.
In that setup, 250 poisoned documents were enough to backdoor models ranging from 600 million to 13 billion parameters. The trigger caused the tested models to produce gibberish. Anthropic explicitly describes the experiment as narrow and says it remains unclear whether the same pattern holds for larger models or more complex behaviours.
Real research. Different commercial conclusion.
The experiment is important security research. It is not a proven content marketing formula.
Known250 deliberately poisoned documents worked in the tested backdoor setup
Not proven250 normal brand articles do not have a proven ability to create permanent AI memory
Not guaranteedPublishing a page does not guarantee future training inclusion or recommendation preference
Useful lessonBuild a stronger body of credible evidence rather than chasing an arbitrary content count
250 generic articles could still achieve very little.
Google’s current generative AI guidance explicitly recommends valuable, unique and non commodity content. One substantial piece of original research can support search, AI answers, PR, social activity, email, sales material and independent references because it contains something worth using.
Google’s current generative AI guidance explicitly recommends valuable, unique and non commodity content. One substantial piece of original research can support search, AI answers, PR, social activity, email, sales material and independent references because it contains something worth using.
What actually helps AI visibility and what is mostly folklore?
Drag or swipe the table sideways to compare each claim.
| Action or claim | Evidence position | Confidence | Priority |
|---|---|---|---|
| Make priority pages crawlable and indexable | Foundational for Google Search. OpenAI says OAI SearchBot access helps ChatGPT discover, surface and clearly cite content. | Known | High |
| Publish useful original evidence | Google explicitly recommends useful, unique, expert led and non commodity content. | Known | High |
| Use clear headings, definitions and textual information | Helps people navigate and gives search systems clearer material to interpret. | Known / supported | High |
| Use keyword and query research | Helps establish demand, market language, commercial pages and customer questions. | Known | High |
| Make decision factors and proof explicit | Improves customer evaluation and provides clearer factual relationships for retrieval based systems. | Supported | High |
| Build genuine external authority | Reviews, credible coverage, customer references and professional sources strengthen the wider evidence environment. | Supported | High |
| Create llms.txt to improve Google AI visibility | Google says it does not use llms.txt for Search. | Known false | No Google benefit |
| Break every article into tiny AI chunks | Google says artificial chunking is not required. | Known false | No requirement |
| Add special AEO or GEO schema | Google says there is no special schema required for generative AI Search. | Known false | No requirement |
| Create one page for every prompt wording | Exact prompt variation pages can drift into scaled, low value content rather than useful customer content. | Supported risk | Avoid |
| Manufacture forum or Reddit mentions | Inauthentic mentions create reputation and spam risk rather than genuine authority. | Supported risk | Avoid |
| Publish 250 articles so AI remembers the brand | The claim is not established by the Anthropic poisoning research. | Unsupported | Avoid |
How to test AI recommendation visibility properly
The first rule is not to rely on one generic prompt. A useful test uses realistic customer situations across understand, find, solve and choose, then records recurring patterns across the AI services which matter to the audience.
Does the organisation appear?
Record inclusion and whether the organisation appears consistently across repeated commercially relevant questions.
Which competitors appear?
Identify the practical competitive set rather than assuming the companies you monitor in SEO are the only alternatives being suggested.
How is the organisation described?
Record services, sectors, strengths, weaknesses and terminology associated with the business.
Which sources support the answer?
Track company pages, press, reviews, directories, customer sites, communities and other cited or retrieved evidence.
What proof is missing?
Find where competitors can substantiate an important claim more clearly than the organisation can.
What happens after the citation?
Review destination page relevance, proof, calls to action, conversion friction and follow up.
Prompt testing is directional.
AI answers can vary by platform, wording, date, location, retrieval behaviour and available sources. The purpose is to identify recurring patterns, omissions and evidence gaps. It is not to create a pretend ranking from one answer.
AI answers can vary by platform, wording, date, location, retrieval behaviour and available sources. The purpose is to identify recurring patterns, omissions and evidence gaps. It is not to create a pretend ranking from one answer.
The check and balance
Every important AI visibility conclusion should pass three questions: what is the evidence, how confident are we and what could the finding actually change?
1. Establish the evidenceCommercial data, customer research, behaviour, search, audit evidence, market evidence or AI recommendation evidence.
2. Apply confidenceKnown, supported, assumed or unknown.
3. Identify potential valueDiscovery, trust, conversion, qualified demand, revenue, efficiency or another defined outcome.
4. Protect what already worksDo not damage ranking pages, strong journeys or existing authority in the rush to “optimise for AI”.
5. Separate correlation from causationA cited page is evidence of visibility. It is not proof that the page caused a sale or recommendation.
6. Research platforms separatelyDo not assume Google, ChatGPT and other services use the same signals in the same way.
7. Test before scalingUse focused content, technical or authority work to learn before creating a large programme.
8. Measure first party outcomesSearch Console, analytics, CRM and sales evidence should outrank directional AI vanity metrics.
9. Keep missing evidence visibleUnknown is a valid conclusion when the data is not strong enough.
10. Improve continuouslyKeep what changes something useful, change what does not and continue learning.
Drag or swipe the table sideways to compare example conclusions.
| Example conclusion | Confidence | Reason | Action |
|---|---|---|---|
| SEO remains relevant for Google’s generative AI Search | Known | Direct current Google guidance | Protect SEO fundamentals |
| OAI SearchBot access matters when ChatGPT Search visibility is commercially relevant | Known | Direct current OpenAI guidance | Review crawler access |
| The 191 homepage audits show uneven online presence foundations | Known | Direct WLW benchmark data | Research the limiting area |
| Making proof clearer may improve the information available for AI recommendations | Supported | Consistent with retrieval based discovery but causal impact needs measurement | Test on priority pages |
| One new case study will increase ChatGPT recommendations | Assumed | Plausible but not directly proven | Measure before scaling |
| A published article will enter future model training and become permanent memory | Unknown | Publication alone does not establish future training inclusion | Do not claim it |
Content as commercial infrastructure
A content pillar is not a target number of articles. It is a commercially useful subject area the organisation has a genuine reason to own.
The strongest pillar connects customer demand, keyword and query research, situations, decision factors, commercial pages, supporting content, proof, authority, digital touchpoints and conversion.
Commercial priority and value creating action
Keyword and query demand
Customer situations and intent
Decision factors and objections
Commercial landing page
Supporting content and customer questions
Owned proof and customer proof
Independent authority and expert proof
Search visibility
AI visibility and recommendation research
PR, social, email, referral and partner touchpoints
Conversion route and customer journey
CRM and follow up
First party measurement
Continuous improvement
Work which can safely wait
Original research is particularly powerful because it can serve several jobs at once. The WLW 191 homepage benchmark is an example: it creates original evidence for this thought piece, the online presence audit, future industry studies, sales conversations, PR and AI visibility research.
How AI visibility should be measured
The measurement hierarchy should keep AI visibility in proportion. First party and commercial measures should sit above directional AI scores.
CommercialSales, revenue, profit, pipeline, qualified enquiries, bookings or another value creating outcome.
BehaviourConversion rate, forms, calls, assisted conversion, repeat visits and journey progression.
SearchImpressions, clicks, CTR, commercial query visibility and branded demand.
AIReferral traffic, citations, recommendation patterns and generative visibility where measurable.
CoverageCustomer situations, decision factors, proof, authority and priority topic coverage.
Content production volume should not become the KPI. Publishing 30 articles is activity. Increasing qualified enquiry volume, growing visibility around a priority commercial subject or improving conversion is an outcome.
B2B and consumer facing businesses still share the same fundamentals
B2B heavy sample
82 matched homepage audits across technology, telecommunications, HR, professional services, energy, manufacturing and construction.
69.7SEO
54.1Perf.
51.1Security
75.0UX
62.5Overall
Consumer facing sample
18 matched homepage audits across retail and consumer goods, travel and leisure plus media, entertainment and sport.
68.7SEO
57.0Perf.
53.6Security
77.3UX
64.1Overall
The 1.6 point overall difference does not establish a meaningful winner and the consumer sample is smaller. The useful point is that both B2B and consumer journeys still need to become discoverable, understandable, trustworthy and actionable. The evidence required to achieve that can be very different.
Three benchmark examples. Three different jobs.
Protect the lead
Caterham Cars
One of the strongest genuine homepage captures in the benchmark.
85SEO
75Perf.
93Security
78UX
Composite: 82.8
- The job is not to rebuild a strong position for the sake of it
- Protect what already works
- Research remaining trust, accessibility, product and ownership content opportunities
Move the middle
Ortus Energy
A strong commercial proposition sitting close to the overall benchmark.
89SEO
50Perf.
40Security
67UX
Composite: 61.5
- Retain the funded energy proposition and customer proof
- Research performance, security, structure and accessibility
- Strengthen answerable authority where evidence supports it
Release trapped authority
SMMT
A highly authoritative organisation whose audited homepage mechanics were considerably weaker than the organisation behind them.
50SEO
0Perf.
40Security
65UX
Composite: 38.8
- The organisation already owns substantial automotive data and expertise
- The homepage snapshot did not communicate that strength particularly well
- The job is to let the online presence represent more of the authority already inside the organisation
Case study boundary
These are point in time homepage audit captures. They do not establish how Caterham, Ortus or SMMT perform in AI recommendation systems. The examples show why the work should follow the evidence rather than one generic AI prescription.
These are point in time homepage audit captures. They do not establish how Caterham, Ortus or SMMT perform in AI recommendation systems. The examples show why the work should follow the evidence rather than one generic AI prescription.
What the work should look like
1. Understand the commercial problemDefine what useful outcome needs to change.
2. Research demand and customersUse keywords, queries, situations, customer evidence, behaviour and market research.
3. Review the current online presenceProtect what works and identify where value is being lost.
4. Research AI visibilityTest relevant platforms and realistic recommendation situations where commercially useful.
5. Establish proof and authority gapsSeparate owned, customer, independent, expert and commercial proof.
6. Prioritise the workstreamsRecommend only the activity with a clear role in changing the outcome.
7. Produce useful content and experiencesBuild commercial pages, supporting content, research, proof and customer journeys.
8. Connect the touchpointsSearch, AI, PR, social, email, referral, website, CRM and other relevant interactions.
9. Measure what happensUse first party commercial and behavioural evidence wherever possible.
10. Improve itKeep what changes something useful and stop activity which does not.
Find out what your online presence is really telling people and AI
WLW can follow the evidence across website structure, search demand, AI visibility, recommendation research, content, proof, authority, brand, UX, technical performance, reviews, customer journeys, CRM, analytics and emerging demand.
You do not need every area to need fixing. You need to know which ones matter.
Book a 15 minute requirements callFrequently asked questions
What is AI visibility?
AI visibility describes whether a business, brand, product, page or source appears or contributes when people use generative AI to research, understand, find, solve or choose. It can include citations, descriptions, comparisons and recommendations.
What is AI recommendation visibility?
It is whether an organisation appears when an AI system is asked to recommend, compare or shortlist providers for a realistic customer situation. It should be tested directionally across several questions and relevant services rather than treated as a fixed ranking.
What is AEO?
AEO stands for Answer Engine Optimisation. It is commonly used for work focused on answer based search visibility. Google currently says its generative AI Search experiences remain rooted in core Search ranking and quality systems.
What is GEO?
GEO stands for Generative Engine Optimisation. It can be useful shorthand for work around generative AI visibility but it should not replace keyword research, customer demand, proof, authority, conversion or measurement.
Is SEO still important for Google AI Overviews and AI Mode?
Yes. Google’s current guidance says SEO best practices remain relevant and foundational because its generative AI Search features rely on core Search ranking and quality systems.
How can a website appear in ChatGPT Search?
OpenAI says any public website can appear in ChatGPT Search. Publishers who want content eligible for summaries, snippets and clear citations should not block OAI SearchBot and should ensure their host or CDN allows OpenAI’s published crawler traffic.
Is OAI SearchBot access the same as allowing model training?
No. Search discovery and potential model training are separate processes. OpenAI publishes separate crawler controls. Allowing search crawling should not be described as a way to force future model memory.
Do I need llms.txt for Google AI visibility?
No. Google’s current guidance says Google Search does not use llms.txt. Maintaining one for another system is a separate choice but it does not improve Google Search visibility.
Do I need special schema for AEO or GEO?
No special AI schema is required for Google’s generative AI Search features. Normal supported structured data can still help conventional Search features where it accurately represents visible content.
Should content be broken into tiny chunks for AI?
No. Google explicitly says there is no requirement to break content into tiny pieces for generative AI understanding. Clear sections and headings are useful because they help readers and search systems, not because there is an arbitrary AI chunk size.
Will 250 articles make ChatGPT remember a company?
No evidence establishes that. The 250 figure comes from an Anthropic data poisoning experiment using deliberately malicious documents and a specific backdoor behaviour. It is not a proven brand content threshold.
What kind of content is strongest for AI visibility?
Useful material containing genuine information worth retrieving: original research, first hand expertise, clear definitions, product and service information, customer evidence, case studies, technical data, real comparisons, decision guidance and current facts.
Can reviews and third party sources influence AI answers?
They can form part of the wider information environment and may be surfaced by search and AI systems. Their importance varies by platform and query, so reviews, trade coverage, professional bodies, customer references and genuine community discussion should be researched where commercially relevant.
How should AI visibility be measured?
Start with commercial and first party measures: qualified enquiries, sales, pipeline, customer behaviour, Search Console and analytics. Add AI referral traffic, generative search reporting where available and repeated recommendation testing as supporting evidence.
Sources and methodology
- WLW FUTURE UK Online Presence Benchmark 2026. 191 matched completed homepage audits across 14 industries from a strictly filtered company pool of 1,540 domains with 4 to 200 employees. Read the benchmark.
- WLW FUTURE v2.1 online presence audit and strategic analysis framework. Used to apply evidence hierarchy, confidence levels, commercial prioritisation, proof analysis and AI visibility boundaries.
- WLW FUTURE v2.1 content pillar, search and AI visibility strategy framework. Used to connect commercial demand, keywords and queries, customer situations, understand/find/solve/choose intent, decision factors, proof, authority and conversion.
- Google Search Central: optimising for generative AI features on Google Search.
- Google Search Central: new resource for generative AI optimisation, May 2026.
- OpenAI: Publishers and Developers FAQ.
- OpenAI: Searching the web with ChatGPT.
- Anthropic, UK AI Security Institute and Alan Turing Institute: A small number of samples can poison LLMs of any size.
External research note: Platform guidance describes current product and crawler behaviour. It does not prove that a particular content change will cause a particular AI recommendation. AI recommendation research is therefore treated as directional unless stronger evidence exists.
