Total Graph Authority, How Brands Get Cited by AI.

Ask ChatGPT the same question twice and you will often get two different answers that cite two different sets of sources. Ask Gemini, Perplexity, and Google AI Overviews the same question and the gap gets wider. Large language models assemble each answer fresh from whatever they retrieve and trust at that moment, so the output moves around even when the question does not.

For a brand, that changes what visibility is worth. Turning up in one answer proves very little. Turning up across repeated runs, and across different assistants, is what actually sends you traffic and shortlists you. We call the position that gets you there Total Graph Authority. This post covers what it is and how to build it, with examples you can copy.

We build this for a living, and the pattern below is what we see over and over in the prompt sets we track for clients. None of it depends on a trick. It depends on being present, and consistent, in every place a model goes looking.

What Total Graph Authority means

Total Graph Authority is the position a brand holds when three independent sources of trust all point at the same company and describe it the same way. Your own domain says you are a strong answer. Other people's sites back that up. Social platforms show it in practice. When the three agree, the specific path a model takes through its sources stops mattering, because they all arrive at you.

Most brands do one of these well and hope the rest sorts itself out. They build a good website and wait to be quoted. They chase links and never touch social. They post on TikTok while their own site stays impossible to retrieve cleanly. Each is one entry in a draw the model re-runs on every query. You want an entry in every draw.

Why one strong channel is not enough

Here is one buyer question run three times through the same assistant. The sources it pulls change each time, even though the prompt is identical.

PROMPT  "what is the best project management tool for a small agency"

RUN 1  cited: your-blog.com, g2.com, a youtube review
RUN 2  cited: reddit.com/r/agency, capterra.com, a competitor blog
RUN 3  cited: your-blog.com, a linkedin post, an aggregator "top 10" list

A brand that only owns its own site takes Run 1 and Run 3, then loses Run 2, because Run 2 leaned on Reddit and an aggregator where the brand was nowhere. A brand with Total Graph Authority appears in all three runs, because it had built a presence in every source the model happened to reach.

The three sources line up with the three ways models decide what to trust, and you have to earn each one. The rest of this post takes them in turn, then shows what happens when they overlap.

Source one, first-party authority

First-party authority is everything on your own domain. It is where a model works out what your brand is, and where it lifts clean, quotable answers. Both depth and structure matter. Five thin pages will not make you an authority on a category, and a deep site with no structure will not get retrieved cleanly.

Build educational hubs, not scattered posts

Assistants favour real topical depth from one coherent source. A hub is a pillar page plus a cluster of supporting pages, all interlinked around a single topic. This is the shape of a hub for a made-up logistics tool called Northwind.

/guide/freight-management/                (pillar, definitional)
  /guide/freight-management/ltl-vs-ftl/   (comparison)
  /guide/freight-management/rate-shopping/ (how-to)
  /guide/freight-management/customs-docs/  (checklist + FAQ)
  /guide/freight-management/glossary/      (definitions, entity-dense)

The glossary and the definitional pillar carry most of the citation weight, because they answer the "what is" and "how does" questions assistants field all day, in a format that is easy to lift.

Write blocks a model can quote

A quotable block is a short, self-contained answer a model can take whole, without needing the rest of the page around it. Structure beats prose here. Compare two versions of the same fact.

WEAK   "There are a lot of factors that go into LTL pricing, and
        it can be complicated to understand how carriers work it out."

STRONG "LTL freight is priced on five factors: freight class,
        weight, distance, density, and accessorials. Freight class
        (from 50 to 500) has the largest single impact on the rate."

The strong version gets quoted because it is specific and stands on its own. The weak version gets skipped because there is nothing in it to lift.

Remove ambiguity with structured data

Models read schema to work out who you are and how your pages connect. At a minimum, ship Organization, Article, and FAQ markup. Here is a trimmed Organization block that ties the brand to its real identifiers.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://northwind.example/#org",
  "name": "Northwind Logistics",
  "url": "https://northwind.example",
  "sameAs": [
    "https://www.linkedin.com/company/northwind",
    "https://www.crunchbase.com/organization/northwind",
    "https://en.wikipedia.org/wiki/Northwind_Logistics"
  ],
  "knowsAbout": ["freight management", "LTL shipping", "customs documentation"]
}

The sameAs array does the real work. It tells the model that the Northwind on your site, on LinkedIn, on Crunchbase, and on Wikipedia is one company, which is how signals from different places get credited to a single brand instead of scattering.

Ship an llms.txt

An llms.txt at your domain root is a short file that points AI crawlers at the pages you most want cited. Small effort, useful for retrieval.

# Northwind Logistics

> Freight management software for small and mid-size shippers.

## Core guides
- [Freight management, explained](/guide/freight-management/): the definitional pillar
- [LTL vs FTL](/guide/freight-management/ltl-vs-ftl/): when to use each
- [Freight glossary](/guide/freight-management/glossary/): 120 defined terms

## About
- [Company](/about/): founders, history, coverage

Source two, third-party validation

Third-party validation is other people vouching for you. It carries more weight than most brands expect, because a model trusts corroboration from places it does not think you control. In the prompt sets we track, third-party pages are cited more often than the brand's own domain, so the site you do not own can matter more than the one you do.

Win the comparison queries

A large share of buyer prompts are comparative. "Best X", "X vs Y", "alternatives to Z". Those answers usually come from third-party listicles and comparison pages rather than vendor sites. You cannot publish the neutral "top 10" yourself, but you can earn a place in the ones people already read, and you can hand them a clean, honest comparison they can lift.

A comparison that admits where a rival is stronger gets published and cited. One that pretends you win every row gets ignored, so be honest about the rows you lose.

Show up where the model already looks

A few third-party surfaces get cited far more than their size suggests. Build a real presence on the ones that fit your category.

• Aggregators and marketplaces like G2, Capterra, and category directories. Claim the profile, keep it current, gather reviews.
• Reddit and forums. Assistants quote these constantly for candid opinion. Earn genuine mentions and do not astroturf, since it reads as fake and gets you nothing.
• Wikipedia and Wikidata if you meet the notability bar. A Wikidata item is often what anchors your entity in the knowledge graph.
• Trade and news coverage from digital PR. One original data study can earn dozens of citing links.

Give journalists a number to quote

Editorial links come from stories, and the easiest story to place is a figure nobody else has. A sample angle for Northwind, invented here to show the shape.

ANGLE  "We analysed 40,000 LTL shipments. Freight was misclassified
        on 1 in 6, costing shippers an average of $214 per load."

WHY IT WORKS  a specific, quotable stat that trade press, Reddit,
             and other blogs repeat, and each repetition is a
             third-party citation that credits Northwind as the source.

Seed a genuinely useful forum answer

Not a plug. An answer that helps, and names your approach where it fits.

Q  "How do you stop overpaying on LTL?"

A  "The biggest leak is freight class. Reclassify your top 20 SKUs
    first, since class drives most of the rate. We built Northwind
    partly because doing this by hand missed it every time, but you
    can run the first pass in a spreadsheet."

Source three, social utility

Social has become a citation channel, not only a reach channel. Assistants pull from YouTube, TikTok, and Instagram for one growing class of question in particular, the ones about whether something works and who it works for. "Does X do Y", "is X worth it", "how do I do Z with X". Transcripts and captions get indexed, quoted, and linked.

Target the "does it work for me" questions

These are the questions social content answers better than a definition can, because they want to see the thing in use.

• "Is Northwind good for a 3-person brokerage?"
• "Can Northwind handle cross-border customs docs?"
• "How do I set up rate-shopping in Northwind?"

Write the metadata for the query

A model reads your title, description, and on-screen text before it processes any footage, so write those for the question you want to answer.

YOUTUBE TITLE  "Is Northwind Worth It for a Small Brokerage? (Honest 3-Month Review)"

DESCRIPTION    "We tested Northwind for LTL rate-shopping and customs docs.
                What worked, what did not, and who it is actually for.
                Timestamps: 00:00 setup, 02:10 rate-shopping, 05:40 customs."

ON-SCREEN TEXT "Best for: small shippers doing 20 to 200 loads a month."

That description answers who the tool is for, what it does, and hands the model timestamped structure to quote. A short vertical cut of the same video, with the "best for" line as on-screen text, carries it onto the platforms younger buyers search first.

Where the three overlap

Put the three together for one query and the outcome shifts. In the first version Northwind owns only its site. In the second it has built all three sources around the same company and the same claims.

PROMPT  "best freight tool for a small brokerage doing customs"

BEFORE (first-party only)
  RUN 1  cited northwind.example  (in)
  RUN 2  cited a reddit thread, an aggregator  (out)
  RUN 3  cited a youtube review  (out)
  result: named in 1 of 3 runs

AFTER (Total Graph Authority)
  RUN 1  cited northwind.example + a g2 profile  (in)
  RUN 2  cited a reddit answer naming Northwind + an aggregator listing it  (in)
  RUN 3  cited a youtube eligibility review + the northwind glossary  (in)
  result: named in 3 of 3 runs, described the same way each time

The model behaves the same in both versions. The difference is Northwind. Every path the retrieval can take now reaches a source that names the brand, and names it consistently. That is why the approach holds up under answers that keep changing. It does not need the model to be predictable. It needs you to be present wherever the model looks.

How to build it, in order

Order matters, because each layer makes the next one work better. This is the sequence we run. It is not a fixed timeline, since how far each step goes depends on where your authority starts.

1. Fix the entity first. One brand name used consistently, one canonical site, Organization schema with a full sameAs, a Wikidata item if you qualify. Everything after this credits back to it.
2. Build the first-party hub. A pillar, a cluster, and a glossary, all quotable and interlinked. Ship llms.txt.
3. Earn third-party corroboration. Claim aggregator profiles, publish one original data study, arm the comparison pages, seed real forum answers.
4. Add social utility. Utility and eligibility videos with query-shaped metadata on YouTube, then TikTok and Instagram.
5. Measure, then compound. Track what moves, put more into the sources that turn into citations, and revisit the plan each quarter.

How to tell it is working

AI visibility is measurable, so measure it. We track four numbers across a named set of prompts, and we publish exactly how on our methodology page. In short.

• Mention frequency. The share of tracked prompts where the brand is named at all.
• Citation share. Of the sources an assistant cites, how many are yours or vouch for you, prompt by prompt, against named competitors.
• Sentiment. How favourably each mention frames you.
• AI-attributed traffic. Clicks and leads that start on an AI surface.

Start with a small, stable prompt set, grouped by intent, so a trend is readable rather than noise.

INTENT        SAMPLE PROMPT
branded       "what is Northwind Logistics"
category      "what is the best freight tool for small shippers"
comparison    "Northwind vs Alternative A"
problem       "how do I stop overpaying on LTL"
eligibility   "is Northwind good for a 3-person brokerage"

Run that set weekly across ChatGPT, Gemini, Perplexity, and Google AI Overviews. The four numbers turn a vague sense of "are we showing up in AI" into a line you can act on.

The point

When answers keep changing run to run, betting on a single source is weak. Your own content makes you retrievable. Mentions on sites the model already trusts make you credible. Video and social show that you work in practice. The three have to agree on the same brand and the same facts, or the model gets a muddled picture and reaches for someone clearer.

Get all three pointing at one company and you reach Total Graph Authority. From there you stay in the answer whichever way a given query resolves, which is the closest thing to a durable position in AI search right now.

If you want to see where your brand sits across the three sources today, we run a free audit against your named prompts and three competitors. The strategy view of this framework lives on our AI Search page.