[[FIG:kb-0285]]

The four source categories

1. The training corpus

The body of text the model was built on, fixed at its training cutoff: web pages, news archives, books, Wikipedia, and structured datasets. This is the model’s baseline understanding of an entity. The knowledge is baked into the model’s weights, so it persists even when no live retrieval runs, but it ages, and model providers do not disclose what gets included or upweighted during training.

2. Retrieval-augmented generation (RAG)

Live web pages fetched at the moment the user asks a question, rather than the training corpus alone. Engines including Perplexity, ChatGPT Search, and Google AI Overviews work this way. What ranks on the live web at query time shapes the answer in real time, so a brand’s current web presence matters independently of what the model learned at training.

3. Structured knowledge bases

Databases of entity facts, Wikidata being the most prominent, that act as a central hub linking an entity’s representations across sources and languages. A single structured record can surface across many engines and locales, so entity accuracy in these databases is its own reputation lever.

4. User-generated content (UGC)

Reddit threads, YouTube, and platform-specific forums. This category is now a mainstream AI citation source: a 2026 analysis of roughly 30 million sources found Reddit the most-cited domain in AI-generated answers, followed by YouTube and LinkedIn. A separate 2026 report found the share of AI citations attributed to social media rose from late 2025 through early 2026, topping 9% of all citations. Classic SEO programs that ignored these channels leave a large gap in the AI-visible brand narrative.

Why a Google-only program misses most of this

The training corpus is set at the model’s cutoff, and no single search campaign moves it in real time. The retrieval layer pulls from the live web at query time. Structured knowledge sits in entity databases. And the user-generated layer lives in communities, Reddit, YouTube, forums, that classic SEO has tended to ignore. Influencing what AI engines say about a brand means working across all four categories.

The post What data sources do AI models use to answer questions about brands? appeared first on Five Blocks.

[[FIG:kb-0294]]

Channel 1, training data

Wikipedia was part of the foundational training corpus for the leading AI models. BERT, one of the architectures that influenced most modern language models, was trained on roughly 2.5 billion words of English Wikipedia alongside book text. GPT-3 included approximately 3 billion Wikipedia tokens in its training mix. Because this training happens before any specific query is asked, the model already has a baseline description of most entities with Wikipedia articles before it ever touches the live web.

Channel 2, live retrieval

Retrieval-equipped engines, those that issue live web searches at query time, consistently cite Wikipedia among their highest-weighted sources. Semrush data from a multi-month citation study found that Reddit and Wikipedia remained ChatGPT’s two most-cited domains. Profound’s analysis of citation patterns across AI platforms found that Wikipedia is ChatGPT’s most cited single source, accounting for around 7.8% of total citations in the measured period. When an engine fetches pages to build an answer, the Wikipedia article for the named entity is typically one of the first pages it pulls.

Channel 3, structured entity data

Wikipedia is closely tied to Wikidata, which acts as a central hub linking all language versions of an article to one underlying entity record. Wikidata’s sitelinks allow easy navigation between language versions using the linked Wikidata item as a reference point, which is part of why consistent facts surface across language editions. Google’s Knowledge Graph, which powers Knowledge Panels and feeds Gemini and AI Overviews, draws on this same layer. So even an engine that is not reading the Wikipedia article text may be retrieving entity facts, founding date, headquarters, leadership, category, that originate from the same article’s infobox.

Why the verbatim feel happens, and what the evidence supports

AI text that closely tracks specific Wikipedia phrasing is the output of all three channels reinforcing the same source. The training weight, the live-retrieval preference, and the entity-data layer all point at the same article, so the synthesis follows it. How closely any specific model echoes the exact article wording varies by model architecture and query type, and no published measurement has quantified a verbatim match rate. What the citation data does establish is that Wikipedia is one of the heaviest-weighted inputs. If the article contains an error, an awkward framing, or an outdated description, that version of the company is what the AI tends to surface.

What to do about it

Because all three channels trace back to the same article, improving the article is one of the highest-leverage interventions in an AI reputation program. The work runs through proper disclosed conflict-of-interest channels: edit requests on the Talk page citing reliable secondary sources, sourcing improvements, and neutral-point-of-view maintenance. The aim is not a flattering article but an accurate, balanced, well-sourced one, because those are the qualities that keep the article credible enough for the engines to keep relying on it.

The post Why does ChatGPT seem to pull my company’s Wikipedia article verbatim when I ask about us? appeared first on Five Blocks.

How coverage maps to the answer

• Well-documented controversy, when a story has been covered by high-authority outlets such as Reuters, Bloomberg, the Financial Times, and the New York Times, AI engines surface it consistently, often quoting or paraphrasing that coverage.
• Contested or thinly-sourced controversy, when something has only appeared in lower-authority outlets or remains disputed, the engines weight it less heavily or present several framings rather than a single verdict.

This comes down to how the engines work: search and AI engines weight sources by credibility, so citation by credible, authoritative, independent outlets shapes AI answers more than additional owned pages do.

[[FIG:kb-0302]]

Why the model is the wrong place to intervene

You cannot edit or manipulate AI model outputs directly. Influence comes from shaping the sources, entity signals, authoritative content, structured data, that the models draw on. The leverage on a controversy sits in the source ecosystem, not inside the engine.

What that work looks like

1. Provide accurate context through Wikipedia, which is built from reliable secondary sourcing such as media coverage.
2. Make sure the brand’s official response is visible and well-structured on owned properties.
3. Work with credible third-party sources where appropriate to establish the accurate record.
4. Track with AIQ™ to see how the source weighting changes over time.

The goal is not to make the engines say nothing. It is to make sure what they say is accurate, complete, and in the right context.

The post How do AI models handle controversial or negative information about brands? appeared first on Five Blocks.

[[FIG:kb-0287]]

Five source levers, work these, not the model

1. Wikipedia improvement. For most notable companies and individuals, the Wikipedia article is the anchor the engines keep returning to; they paraphrase it, quote it, and treat it as the reference. Improving the article through Talk-page requests and disclosed conflict-of-interest editing is usually the single highest-leverage step. Retrieval-heavy engines can reflect accepted edits within days to weeks; training-anchored engines respond over months as the broader ecosystem catches up.
2. Knowledge Graph and Wikidata corrections. When a structured-data error is driving the wrong answer, a wrong founding date, a misclassified entity type, a broken sameAs link, fixing it at source pushes the correction across Google’s Knowledge Panel and into engines that query structured data directly. Wikidata links entity records across language editions, so one correction carries across multilingual surfaces.
3. Owned content strengthening. FAQ pages, executive bios, and pillar content written in clear, extractable language give engines first-party material to draw on. Schema markup on owned properties helps engines attach that content to the right entity. Third-party authority still outweighs owned pages in most engine weighting, so this lever works best alongside steps 4 and 5.
4. Third-party coverage in trusted outlets. AI engines weight authoritative press heavily. Placements in outlets the engines trust, the ones that rank independently for relevant queries, enter the source pool and move the narrative more reliably than additional owned pages. Retrieval-equipped engines can pick up new coverage in the same news cycle; the effect on training-anchored engines builds over months.
5. Wikidata structured corrections. Fixing factual errors in Wikidata entries, names, dates, relationships, identifiers, closes a low-visibility gap that produces stubbornly persistent wrong answers across multiple engines. Wikidata entries feed the Knowledge Graph and the sameAs infrastructure engines use to resolve entities.

What ‘visible progress’ looks like, and when

Progress comes in two phases. Retrieval-heavy engines (Perplexity, Google AI Overviews, ChatGPT Search) can reflect source-layer changes within days to weeks of an authoritative edit or a credible new placement going live. Engines anchored more to their training data respond over months, as the broader web absorbs and republishes the improved source material. A full, durable shift across the major AI engines, ChatGPT, Copilot, Gemini, AI Overview, Perplexity, Grok, Claude, and Google AI Mode, usually takes a six-to-twelve-month engagement. AIQ™ monitors the eight engines it currently tracks continuously, so the trajectory shows up in monthly reporting well before the engagement ends.

The post Can you influence what AI says about your company? appeared first on Five Blocks.

[[FIG:kb-0286]]

Three audiences the AI narrative reaches first

Follow the three audiences most likely to open an AI engine at the start of their research:

• Journalists research subjects before making contact, and the AI synthesis is now part of that preparation. The narrative they receive frames the angle they bring to everything else they read.
• Investors and allocators prompt AI engines about prospective investments before formal diligence begins. The story the engine tells shapes which questions get asked and which risks surface first.
• Senior candidates research employers in ChatGPT and similar tools before they apply or accept an offer. The narrative about an employer decides whether a candidate arrives already favorable or already skeptical.

In each case the narrative shapes how the reader treats everything else. A clear positive framing makes supporting evidence feel confirmatory; a thin or negative framing makes even strong evidence feel insufficient.

What a positive versus negative narrative looks like in practice

Take one company at two different moments in its public record. When the authoritative source ecosystem is dominated by earned media covering product milestones and leadership credibility, the AI synthesis tends to lead with category leadership, competitive differentiation, and growth. When that same ecosystem is instead dominated by a controversy cluster, regulatory action, a leadership departure, or sustained critical coverage in major outlets, the engines put the controversy first, and even good financial results land inside a risk frame. The facts of the business have not changed; what changed is which sources the engine weights, and therefore what story it tells. AI engines reflect the source ecosystem on controversies, surfacing consistently what major outlets have covered consistently, so the narrative a reader receives is an output of the weighted source record, not a balanced editorial judgment.

How narrative quality is measured

Measuring an AI narrative means polling the engines directly with a defined set of prompts, then reading the responses on three dimensions:

• Inclusion. Is the brand mentioned at all in response to category and competitive queries, or is it absent from the synthesis?
• Framing. When it appears, what themes, adjectives, and associations surround it? Is the lead claim about innovation and leadership, or about risk and controversy?
• Consistency. Does the narrative hold across engines and prompt phrasings, or does it fragment engine to engine, which signals a thin or contested source record?

Tools that track AI visibility report brand citation rates and the text of responses over time, so you can see when a narrative shifts and which engine is driving the change. AIQ polls eight major engines (ChatGPT, Copilot, Gemini, AI Overview, Perplexity, Grok, Claude, and Google AI Mode); category tools such as Profound and Peec do the same.

Why narrative is the new primary unit

The SEO era centered on rankings, meaning which result appeared first for a keyword. The PR era centered on impressions, meaning how many people saw a placement. The AI era centers on narrative because the engine collapses the source layer into a single synthesized answer. You no longer compete for position in a list; you compete for inclusion and framing in a synthesis. That is a different problem, and it has to be tracked at the narrative layer, not the link layer.

The post What is an AI narrative and why does it matter? appeared first on Five Blocks.

[[FIG:kb-0299]]

Speed by engine type and change type

Note on the tonal-shift row: the distinction between factual corrections moving faster than tonal shifts is practitioner framing derived from the supported source-synthesis mechanism; it is not an independently benchmarked statistic.

Why the two-clock model matters

• Retrieval-first engines issue a live web search at query time. Perplexity describes its own mechanism as providing “real-time access to ranked web search results from a continuously refreshed index.” Google documents AI Overviews as “relying on our core Search ranking systems to retrieve relevant, up-to-date web pages”, grounding responses in the live index rather than static training data. A source change that clears Google’s index can reach these engines within the same update cycle.
• Training-baseline engines reflect no change until the model is retrained or fine-tuned. OpenAI’s published model documentation, for example, records knowledge cutoffs in specific calendar months, which means a model in active deployment may draw on a corpus captured six months or more earlier. Improvements made in the live web ecosystem accumulate and feed the next training cycle rather than appearing on the next query.
• Mixed-engine reality: a reputation program runs across both types at once. Retrieval-first wins appear earliest; training-baseline engines take longer to reflect the same source improvements. AIQ tracks the trajectory across all monitored engines, so clients see direction and rate of change in their monthly reporting rather than waiting in the dark for a final answer.

The post How quickly are AI models’ perceptions of a brand likely to change? appeared first on Five Blocks.

The ranking inputs

Recency

Because Perplexity reads the live web, newly published authoritative content can enter the ranked set at query time rather than waiting for a retraining cycle. A recent, authoritative article on a topic tends to surface ahead of older coverage on the same subject.

Domain authority

A domain’s established reputation is one of the stronger inputs. Well-cited domains, government, academic, major news, and reference sites such as Wikipedia, tend to surface consistently; thin, low-authority pages appear far less often.

Topical relevance

A page has to match the question. A specialist source that covers a niche topic in depth can outrank a higher-profile generalist outlet that addresses it only in passing.

Link and citation patterns

How pages reference one another, the citation graph around a source, feeds into how the returned pages are ranked.

Structured data

Machine-readable signals such as schema markup help an engine understand what a page asserts and classify its content, which makes retrieval and matching cleaner.

The inline-citation verification layer

Once Perplexity writes an answer, it shows the sources it drew on as inline citations. A reader can click through and see exactly which pages the answer was built from. That transparency is also why Perplexity is comparatively easy to monitor and influence: you can see which sources hold the citation slots and direct source-layer work at those specific pages.

What this means for your program

Because Perplexity is retrieval-first, it is often the first place you see source-layer work take effect. Improvements to authoritative pages, a strengthened Wikipedia paragraph, a well-structured corporate bio, recent third-party coverage, can shift what Perplexity cites faster than on engines that lean more heavily on a fixed training baseline. Watching which sources appear in Perplexity’s citation slots for your main queries is a direct diagnostic: it shows what is shaping the narrative today and where source-layer investment is most likely to move the answer.

The post How do AI-powered search engines like Perplexity rank and cite sources? appeared first on Five Blocks.

[[FIG:kb-0291]]

Side-by-side comparison

What to do differently for each

Both engines use retrieval-augmented generation, so the same source-layer work helps both: authoritative third-party coverage, a well-maintained Wikipedia article, clean schema markup, and structured entity signals. The differences are at the margin:

• For ChatGPT Search: conversational prompt patterns mean users often phrase queries as full questions (“What do people say about X?” or “Is X reputable?”). Content that directly answers those question forms: FAQ pages, Q&A-structured editorial, clearly attributed quotes from credible sources, tends to surface in conversational retrieval. ChatGPT also draws on a broad training corpus (books, news, forums, Reddit) in addition to live retrieval, so consistent long-term authoritative coverage across many source types matters alongside freshness.
• For Google AI Overviews: the entry point is a keyword query in Google, so standard E-E-A-T signals, page-level technical factors (crawlability, schema markup, passage-level extractability), and overall index authority apply directly. Google AI Overviews are rooted in Google’s core Search ranking systems, meaning a page already ranking well for a branded keyword is a strong candidate to be synthesized into the Overview. Prioritize pages that are both highly ranked and written in fact-dense, extractable prose.

The post What is the difference between ChatGPT Search and Google AI Overview? appeared first on Five Blocks.

[[FIG:kb-0290]]

The brand-specific hallucination types

The dangerous ones are inventions that sound entirely plausible:

• Fabricated lawsuit, legal action attributed to the company that never happened.
• Fake executive, a name appended to a role the person never held.
• Non-existent product feature, a capability or product that was never shipped.
• Wrong financial detail, a number that matches no filing the company has ever made.

The risk is that each of these is delivered in the same confident tone as a true statement. A downstream reader, a journalist, a candidate, a customer, has no way to tell the fabrication from the fact.

Why prompting the model does not fix it

The engine doesn’t remember what you tell it, and it builds every answer fresh from the sources it trusts. The false claim is usually anchored to one of those sources: a thin or contested page, an older mix-up, sometimes nothing identifiable. Correcting it means working on that source, not arguing with the model.

The fix: working on the sources

1. Identify the anchor. Find what the engine is leaning on to produce the false claim.
2. Strengthen the correct version. Reinforce the accurate account through Wikipedia, owned content, and structured data so the right signal outweighs the wrong one.
3. Track it. Monitor through AIQ to verify the hallucination actually drops out of the engines’ answers.

The post What is an AI hallucination and how does it affect brand reputation? appeared first on Five Blocks.

[[FIG:kb-0300]]

How a Perplexity answer gets built

1. Live web search: each query triggers a real-time search against a continuously refreshed index rather than relying on a training cutoff.
2. Ranking: the returned pages are ranked using Perplexity’s own retrieval logic, weighted toward recency, domain authority, topical relevance, and citation patterns.
3. Synthesis: the model writes an answer drawing from the highest-ranked pages.
4. Inline citations: the sources it used are shown inline, so a reader can click through and verify each claim.

Which pages win the citation slots

Because recency is weighted heavily, a recently published authoritative article often outranks an older one on the same topic. Government, academic, major-news, and Wikipedia domains rank consistently high; thin blog content rarely appears at all.

Why this makes Perplexity an early indicator

The practical consequence is fast feedback. Because Perplexity reads live pages instead of waiting for a retraining cycle, a strong new authoritative article, or an improved Wikipedia paragraph, can shift its answer within days. The same intervention takes longer to surface in engines that lean more heavily on their training-data baseline. Perplexity tends to be the first place you can see that source-layer work is landing.

The post How does Perplexity AI source information about companies and people? appeared first on Five Blocks.