AI Visibility June 20, 2026

How Dattva Works: The Methodology

Most AI visibility tools ask one model. Dattva queries ChatGPT, Perplexity, Gemini, and Claude independently, cross-verifies the results, and tells you exactly what to fix. Here is how the five-stage system works — from diagnostic through ongoing measurement.

When your next enterprise client opens ChatGPT and types the name of your category, does your brand appear? Most companies do not know. And the ones that have checked often find something worse than absence — a competitor showing up in their place, consistently, across every AI model their buyers use.

The follow-up question is always: why? Not in a general sense. Specifically — which part of the setup is wrong, what is the AI actually saying about the brand when it does find it, and what exactly needs to change. That is the question most AI visibility tools do not answer. They show you the gap. They do not close it.

We run a five-stage system to close that gap: diagnostic first, visibility implementation second, agent readiness third, hallucination detection fourth, and ongoing measurement fifth. Each stage depends on the one before it. The order is not arbitrary — it reflects how AI visibility actually breaks down and gets fixed.

1. Diagnostic

Before we touch anything, we measure where you actually stand. We do not start with assumptions. We run the actual buyer queries your prospects type into AI — across ChatGPT, Perplexity, Gemini, and Claude — and record exactly what comes back. Not a sample. Not estimated. Real responses, per platform, per query.

This produces a citation map. It shows which queries your brand appears in, which it does not, and who appears in your place. The platform distinction matters — a brand can appear consistently on Perplexity and be completely invisible on ChatGPT for the same query. We show that gap explicitly, not as an average score across engines.

We score each brand across four dimensions:

  • AI Brand Awareness — how frequently and accurately your brand is named in responses to relevant buyer queries
  • Technical Readiness — whether AI crawlers can access, parse, and correctly categorise your site
  • Narrative Strength — whether what AI says about your brand matches what your brand actually says about itself, across multiple platforms
  • Retrieval Quality — how often your content pages are pulled as source citations in retrieval-based models like Perplexity

Most brands score between 35 and 52 out of 100 before any work begins. The industry benchmark we use is 65. These numbers come from our own diagnostic runs across clients in IT, BFSI, retail, and auto — not a survey, real runs.

The output is a locked baseline. We compare every implementation result against it. It does not get thrown away after the first report.

Honest limitation

Part of the diagnostic involves a human reviewing AI responses and cross-checking them against the brand's actual website. It is not fully automated. That review takes 24 to 48 hours. We say this because it is the part that makes the output accurate.

2. Visibility Implementation

This is where the gap gets closed. It requires two layers working together — technical and content. Neither alone is sufficient.

The technical layer

The most common finding in our audits is also the most invisible to the site owner: AI crawlers are being blocked.

GPTBot, ClaudeBot, and PerplexityBot are frequently denied access through robots.txt rules or Cloudflare security defaults that were set up for human traffic, not AI crawlers. The content is indexed on Google. The AI cannot see it at all. We fix this with explicit Allow directives for each AI crawler and verify that critical content renders server-side in HTML — not via JavaScript that crawlers cannot execute.

Schema markup is the second technical priority. We implement three types, in sequence, because each does a different job:

Organisation schema tells AI models what the brand is, what it does, and how it connects to verified external entities — Wikidata, LinkedIn, Crunchbase. Without this, Gemini especially tends to generate its own description of a brand from whatever it finds first, which may be inaccurate.

Article schema marks a specific page as authoritative content worth citing, not just a web page. It specifies authorship, publication date, and topic category.

FAQPage schema tells AI models exactly which question a piece of content answers, word for word. AI does not read an entire page the way a human does. It extracts from the part of the page that most directly answers a query. FAQ schema removes the ambiguity of which part that is.

Entity consistency is the third technical fix. AI models — Gemini in particular — cross-reference brand descriptions across a brand's own site, LinkedIn, Crunchbase, and directory listings. If those descriptions differ, the model treats the brand as ambiguous and either hedges or avoids the citation. Fixing this is not cosmetic. It is foundational.

The content layer

We write content for AI extraction, not for keyword ranking. These are different tasks.

Every piece of citation-native content we produce leads with a 40 to 60 word direct answer block. AI models extract from the first clean, complete paragraph of a page. A page that buries its answer in paragraph four will be passed over in favour of one that answers the question immediately. This page follows the same rule.

The hardest thing to accept, and the most important: AI models treat a brand's own website as marketing copy. Third-party sources — G2, Clutch, Wikidata entries, editorial coverage — are treated as independent evidence. A brand with strong internal content but weak external citation presence will consistently lose citations to a competitor with average content but strong third-party presence. Building that external footprint is part of what we do, not an optional add-on.

Results vary by category, starting score, and competitive density. We do not claim a guaranteed outcome. We claim a documented, verifiable process.

3. Agent Readiness for Websites

Most GEO work stops at getting cited in AI chat responses. That is not the full picture any more.

AI agents do not just read a page and summarise it. They navigate websites, fill in forms, extract structured data, and take actions on a user's behalf. A site that is readable by a human and citable by a chat model can still be completely unusable by an agent if its structure is unclear or key actions are buried in JavaScript an agent cannot parse.

The share of AI-driven web interactions that involve agents — rather than passive question-answering — is growing. ChatGPT now sends 3.6 times more crawling requests to websites than Googlebot does. (Source: Alli AI via Search Engine Journal, 2026)

What agent readiness means practically: clean semantic HTML throughout the site, forms and calls-to-action that are properly labelled without requiring JavaScript rendering, structured data that exposes what a page allows a user or agent to do, and no critical content hidden behind dynamic loading.

To be direct about scope: this is infrastructure work, not deep agent-action integration. We are not building custom agent workflows. We are ensuring the site's structure does not become a barrier as agent-based browsing becomes a larger share of how buyers interact with the web. The brands that do this now will not need to retrofit it later.

4. Hallucination Engine

AI models sometimes generate claims about a brand that do not match what the brand actually says about itself. This happens because models are trained on a mix of sources, some of which may be outdated, incorrect, or about a different company with a similar name. There is no alert system. The error sits in every future AI response until something changes it.

The errors we find most often in our audits: wrong product category, outdated information that persists after a company pivot, leadership names that belong to a different company, and features incorrectly attributed to a competitor. None put there deliberately. All accumulating quietly from inconsistent training data.

Our detection process: we run a brand's name, core facts, and key product claims through each AI platform on a recurring basis and check the response against a verified fact set built from the brand's own site and approved materials. When a discrepancy appears, it is flagged with the specific claim, the platform it appeared on, and a recommended correction path.

What we do when something is found depends on the type of error. Schema and Wikidata corrections address the structured data the model may use as ground truth. For persistent errors from editorial sources, we work with the brand on publishing a clear correction page that meets AI citation standards.

Timeline is honest: not instant. Perplexity and other retrieval-based models reflect corrections within days because they pull live sources. Models that rely on training data take longer — corrections propagate when the next training cycle incorporates updated sources. We cannot tell you exactly when because that is not within our control.

We cannot guarantee removal of a hallucination from every model. We can correct the authoritative sources models draw from, which significantly improves accuracy on retrieval-based models. For training-data models, we improve the odds. We will not claim more than that.

This is not AI reputation management or sentiment monitoring. It is specifically about factual accuracy in AI-generated answers. The tools, the process, and the outcome are different.

5. Why This Compounds

This is not a one-time fix.

The brands that stay cited are the ones with citation presence across multiple source types — schema, direct content citations, external directory presence, editorial mentions. When one source type loses weight, the others carry it. This is why external citation building is not a one-month task.

The diagnostic baseline we set at the start gets re-measured. We compare month-one scores against month-three and month-six scores, using the same query set against the same platforms. Citations earned through schema and well-structured content tend to hold and compound because each new citation reinforces the brand's entity in AI's understanding of its category.

What this means practically: ongoing visibility requires ongoing measurement. We re-scan on a defined schedule, flag score drops above a threshold, and identify new prompts where competitors are gaining ground. The goal is not to hand over a report and walk away. It is that your diagnostic score six months from now is measurably higher than it was at the start — verified by typing the same queries into ChatGPT yourself, without logging into a dashboard.

FAQ

What is the first thing you do for a new client? +
We run the diagnostic before anything else. We do not start with assumptions about where a brand is visible or invisible — we run the actual buyer queries across all four platforms and record what comes back.
How is this different from an SEO audit? +
An SEO audit measures Google ranking signals. Our diagnostic measures what AI models say about a brand, which queries they cite it for, and which competitors they recommend instead. A brand can rank on page one of Google and be invisible in every AI answer.
What does the citation map show? +
It shows which buyer queries your brand appears in across ChatGPT, Perplexity, Gemini, and Claude — and which ones it does not. For each query where you do not appear, it records which competitor does. The output is per-platform, not an average.
Why does entity consistency matter? +
AI models cross-reference how a brand is described across multiple sources. If your site says one thing and your LinkedIn profile says something slightly different, the model treats the brand as ambiguous. Making descriptions consistent across your site, directories, and Wikidata is one of the highest-impact technical changes we make.
What is a direct answer block? +
A 40 to 60 word paragraph at the start of a page that directly answers the query the page is targeting. AI models extract from the first clean, complete answer they find. Pages that answer the question immediately get cited more often than pages that build to the answer.
Why does third-party citation matter more than my own site? +
AI models treat brand-owned content as inherently biased. Third-party sources — G2 reviews, Clutch listings, Wikidata entries, editorial coverage — are treated as independent evidence. A competitor with weaker internal content but stronger external citation presence will outperform a brand with excellent internal content but no third-party footprint.
What happens when you find a hallucination? +
We identify the specific claim, the platform it appeared on, and the source the model is likely drawing from. Then we correct the authoritative structured data — schema, Wikidata — the model uses as ground truth. For retrieval-based models, corrections typically propagate within days. For training-data models, the timeline depends on the model's update cycle.
Can you guarantee a hallucination will be removed? +
No. We correct the authoritative sources and significantly improve accuracy on retrieval-based models. For models reliant on training data, we improve the odds but cannot control the timeline.
What does agent readiness mean? +
It means the site's structure is usable by AI agents, not just readable by chat models. Clean semantic HTML, properly labelled forms and calls-to-action, structured data exposing what a page allows a user to do, and no critical content buried in JavaScript agents cannot parse.
Why is ongoing monitoring necessary? +
AI citation patterns change without warning. Reddit's share of ChatGPT responses dropped from over 14% to under 2% in weeks. We re-scan on a defined schedule and compare against the original baseline to detect these shifts before they become significant visibility losses.
What if I want help implementing? +
The free diagnostic gives you a score and five prioritised fixes. For the full blueprint with copy-paste code, content briefs, and implementation support, we offer Starter, Growth, and Enterprise plans scoped to what your team actually needs. Every plan starts with the free diagnostic. No long-term contract. Cancel anytime.

Sources

  • Ahrefs — "Best X" listicles account for 43.8% of ChatGPT cited page types: ahrefs.com/blog/best-lists-research/
  • Alli AI via Search Engine Journal — ChatGPT sends 3.6x more crawler requests than Googlebot: searchenginejournal.com
  • Previsible via Search Engine Land — LLM referral traffic up 527% in five months: searchengineland.com
  • Gartner — 25% projected drop in traditional search traffic by 2026: gartner.com
  • Promptwatch via Motley Fool — Reddit's ChatGPT citation share dropped from 14%+ to under 2% (September 2025): fool.com
  • Dattva internal diagnostic data — 35–52 average baseline score, industry benchmark 65 (IT / BFSI / retail / auto client runs)

Run the free diagnostic at dattva.com. No signup. No credit card. Takes about ten seconds.

See where your brand stands across all four AI engines

Get your free AI Visibility Score — real responses from ChatGPT, Perplexity, Gemini, and Claude, with five prioritised fixes at no cost.