Total Graph Authority explained to a five-year-old

A new bakso cart opens at the end of your street, and nobody has heard of it yet. When a visitor asks the neighbours where to eat tonight, the cart does not come up, because the neighbours can only recommend what they already know and trust.

AI assistants such as ChatGPT, Gemini and Google's AI Mode behave in a similar way when someone asks which brand to buy. They answer from what they can find and what they have reason to believe. Total Graph Authority is our name for the position a brand holds when what it says about itself, what other people say about it and how real customers behave all point to the same brand and describe it the same way.

The idea sounds technical, and most explanations of it are. A five-year-old can still follow it if the brand is a bakso cart and the AI assistant is a neighbour, so the cart is where the explanation stays.

The helpful neighbour is the AI assistant

Every street has one neighbour whom everyone asks for advice. Visitors ask her where to eat, where to fix a tyre and which tailor is honest. She does not cook the bakso herself, and she has no reason to favour one seller over another. Her job is to listen to the street and pass on what she believes is true.

She knows things in two ways. Some of it she has carried in her head for years, such as the famous cart near the mosque that has been there since she was a child. The rest she finds out on the day, by walking down the street and asking around before she answers. AI assistants work with the same two stores. One is what the model learned when it was trained, and the other is what it looks up on the web at the moment of the question.

She also answers a little differently each time. Ask her the same question on ten evenings and she will not name the same three places on every one of them. A study by SparkToro and Gumshoe put the same questions to AI tools nearly 3,000 times and found that the list of recommended brands almost never repeated.

That changes what the cart should aim for. Being named every single evening is out of reach for anyone on the street. Being named on most evenings is possible, and it depends on how much the neighbour has heard about the cart and whether everything she has heard agrees.

How the neighbour puts an answer together

When a visitor asks where to eat bakso, the neighbour does a handful of things very quickly. Seeing them one at a time shows where a new cart can drop out long before the answer is spoken.

1. She works out what the visitor is asking, which here is a place to eat bakso nearby tonight.
2. She recalls the carts she already knows and walks down the street to see who else is there.
3. She reads the signs and menus she can reach, and listens to what people on the street are saying.
4. She decides which carts are worth naming, based on how many people mention each one and whether their stories match.
5. She gives her answer and often says who told her.
6. Over the following weeks, what visitors do with her advice shapes what the street talks about next.

A cart can fail at any of these steps. If the neighbour does not understand that Pak Darto sells bakso, he is never considered for a bakso question. If she cannot reach his alley, she never reads his menu. If she reaches it and nobody else has ever mentioned him, she has no reason to pick him over the cart near the mosque.

This is the main reason a brand with a good website can still be missing from AI answers. The website only covers the part where she reads the sign and the menu. Everything else happens on the rest of the street, and a brand that works on one step alone leaves the other five to chance.

Three kinds of proof the neighbour listens to

The first kind of proof is what the cart says about itself. The name is painted on the side, the menu hangs where people can read it and the price board is honest. For a business this is the website, and we call it first-party authority. A cart with no name and no menu gives the neighbour nothing to repeat, which is why what AI knows about a brand is decided first on its own domain.

The second kind is what other people say. The warung owner next door mentions the cart, a food blogger writes about it and the ojek drivers send hungry passengers there. Their words count for more than the seller's own, because they gain nothing by praising him. For a business these are review sites, media coverage and comparison articles, and we call it third-party validation. Most of what AI cites about a brand lives on sites the brand does not control.

The third kind is what customers actually do. There is a queue at seven, children ask for the cart by name and people share photos of their bowls. We call this social utility. It is the hardest proof to fake and the slowest to build, since it only appears when people have eaten there and liked it.

What matters most is that the three agree. When the sign says Bakso Pak Darto, the warung owner says Bakso Pak Darto and the children in the queue say Bakso Pak Darto, the neighbour hears one clear story from three directions. When the sign says one name, the blogger uses another and nobody is queuing, she hears three weak stories about what might be three different carts, and she recommends someone else.

Five things the street remembers about the cart

The three kinds of proof leave traces behind, and we sort those traces into five graphs. A graph here is only a map of what is connected to what. Each one has a plain version on the street and a matching version for a brand online.

The five are covered in technical detail in a separate article on the graphs underneath. What follows is the street version of each, with the mistake we see most often.

The name everyone uses

If half the street says Pak Darto and the other half says Bakso Solo, the neighbour cannot be sure they are talking about the same cart, and every recommendation is split in two. She may also confuse him with another Pak Darto who sells soto three streets away. The entity graph is the record of who the cart is, and it only works when the name, the description and the basic facts are identical wherever she looks.

For a brand the usual problem is drift. The website uses one company name, the LinkedIn page uses an older one, a directory lists a retired product name and the Google Business Profile puts the company in the wrong category. Each difference is small, and together they make the brand look like several half-known companies. Our work starts here for that reason, because every later effort is credited to a name, and the credit is lost when the name is unclear.

The menu that answers the question

A visitor asks whether the bakso is halal, how much a bowl costs and what time the cart opens. A menu that answers those questions in plain words lets the neighbour repeat the answer straight away. A menu that only says the bakso is delicious and made with love gives her nothing she can use. The content graph is everything the cart has written down, and its value depends on whether it answers what people really ask.

Business websites often make the second kind of menu. Pages describe the company in general terms and leave out prices, comparisons, steps and numbers. A study by Mahe Chen and Nick Koudas, What Drives AI Citations?, found that within one website the pages whose titles and paragraphs closely matched the questions people ask were cited more often. A page that states the buyer's question in its title and answers it in the first lines does the job of the good menu.

The signs that point the way

Hand-painted arrows on the corner, a mention on the warung's noticeboard and a line in the neighbourhood newsletter all help people find the alley. Online these are links from other websites. They have been central to search engines for more than twenty years, and they still help a brand's pages get found and indexed.

Signs matter less to the neighbour than most people expect. An Ahrefs analysis of AI Overviews found that how often a brand was mentioned across the web related to its visibility far more strongly than how many links pointed to it, at 0.664 against 0.218. An arrow tells her where the cart is. It does not tell her whether the bakso is any good, and that is why the link graph carries fewer points in our scorecard than it would in a traditional SEO audit.

The people who say the name

This is the graph that moves the neighbour most. When the warung owner, the food blogger and a parent at the school gate all say that Pak Darto's bakso is worth the walk, she starts repeating it. The citation graph is the record of who names the cart, where they say it and whether they describe it correctly.

Not every mention counts the same. A mention from someone the seller paid is worth nothing to her, and a mention from a stranger on the other side of the city who has never tried the bakso is worth very little. A mention that gets the name wrong can even do harm. We count a mention only when the source is independent, knows the category, describes the brand accurately and speaks about it neutrally or well. Three such mentions outweigh thirty listings in directories nobody reads.

The queue at seven

A queue is the street's way of voting. For a brand it shows up as people searching for the company by name and coming back to it. The interaction graph is the hardest of the five to see from outside, and no published study has isolated its effect on AI recommendations, so we treat it as a confirming signal and give it the fewest points.

It still tells us something useful. When the signs, the menu and the mentions all improve and the queue stays the same length, the work is reaching the neighbour and not yet reaching customers. That gap is worth knowing early, because a recommendation that nobody acts on does not last.

The road to the cart has to be open

None of this helps if the gate to the alley is locked. The neighbour cannot recommend a cart she has no way of reaching, however good the bakso is. In Total Graph Authority this is the retrieval layer, and it sits underneath all five graphs.

For a website the locked gate takes three common forms. A single line in the robots.txt file can block AI crawlers by accident. The main text of a page can be loaded by scripts that some crawlers never run, so they see an empty page. A page can also be missing from the indexes of Google and Bing, which many assistants draw on when they look things up.

The same study put a number on how much the road matters. Which pages the assistant's own search retrieved explained 32% of the variation in citations, and the properties of the page itself explained 2.8%. A beautifully written menu behind a locked gate loses to an average menu on an open street.

We check the road separately from the five graphs. An open gate does not make anyone trust the cart, and a locked one makes every other effort invisible. A brand with strong content and no AI mentions often has a road problem before it has a trust problem.

Shortcuts that do not work on the neighbour

The first shortcut is paying people to praise the cart. The neighbour has lived on the street a long time and can tell a friend's honest tip from a paid one. Online this is paid links, link exchanges and networks of sites that publish anything for a fee. We do not use them, because the mentions they produce are the kind she discounts.

The second shortcut is a bigger name tag. Schema markup is the name tag of a website, a block of code that states the company's name, address and profiles in a form machines can read. It is worth having, since it keeps the name clear. It does not make the neighbour like the bakso. Ahrefs tracked 1,885 pages that added schema and found no rise in AI citations from the markup alone, so we treat it as tidiness and not as a growth tactic.

The third shortcut is shouting. A seller who prints fifty banners saying the same thing has not given the neighbour fifty reasons. Publishing large volumes of similar articles works the same way. One menu that answers ten real questions does more than ten banners that answer none.

How we know whether it is working

We ask the neighbour the same set of questions every week and write down what she says. The questions are the ones real customers ask, such as where to find good bakso nearby, which cart is better for children and where to eat for under a certain price. Each question is asked three times on each assistant, because one answer on one evening proves very little.

The first number we report is the mention rate, which is the share of all those answers that name the cart. A cart named in 23 answers out of 100 has a mention rate of 23%. We also record who else she named, which sources she pointed to and whether she described the cart correctly. Because her answers vary, every figure comes with a margin, and we call a change real only when it moves outside that margin.

The language of the question matters as well. In our own test of Google AI Mode in Indonesia, the brands recommended for the same questions in Indonesian and in English had little in common. The average overlap score was 0.30 on a scale where 1 means identical lists, and the test was small, so the figure is a first indication. A cart that is well known to English-speaking visitors can be unknown to the neighbours who ask in Indonesian, so we ask in the language the customers use.

The order we would fix things for the cart

The name board and the gate come first. Pak Darto paints one name on the cart, uses the same name on the banner and asks the warung owner and the blogger to use it too. He also makes sure the alley gate stays open. For a brand this means one name and one description across the website, the business profiles and the directories, and pages that crawlers can reach and read. Nothing else is worth starting until these two are right.

The menu comes next. It should answer what people really ask, in words a tired customer can read at a glance. After that comes the slow part, which is earning the words of other people. Pak Darto invites the food blogger to try a bowl, answers questions honestly when neighbours ask about his ingredients and gives the newsletter something worth printing, such as how many bowls the street eats in a rainy week. For a brand this is editorial outreach, honest comparison pages and original data that others want to quote.

The queue follows from the bowl, and no amount of signage replaces good bakso. The whole sequence takes months, and the neighbour's memory fades when the street stops talking, so the work is kept up after the first results arrive. Total Graph Authority is Search Agency's position on how brands earn these recommendations, and no assistant publishes how it decides. The full method, with the scorecard, the eight measures and a worked sample, is in the free Total Graph Authority Methodology whitepaper.