Source grader

The source grader is a checklist of the places where an AI model can learn what your brand is. Each row names a type of source, the platforms it covers, the action to take, and further reading.

The list is inspired by Aleyda Solis' Content Prioritization in an AI Search Era, which ranks content types on your own website. The source grader extends that to the sources beyond your own website, where most of the description of your brand is written.

Every source falls into one of three circles, depending on how much say you have over it. The circle decides the kind of action.

Treating a review site like your own website wastes effort. Treating a company register like a review site leaves an easy correction undone.

An about page answers who you are. It does not prove that you can be trusted, that you are worth the price, or that you are the specialist in your field. AI systems look for different evidence for each of those questions.

The research by Steven Wilson-Beales into 4,003 AI answers about news publishers shows the pattern. In 76% of the cases, no about page appeared among the cited sources at all. About pages carried a quarter of the evidence when the question was identity. When the question was value, pricing and subscription pages carried a third. When the question was trust, standards pages, corrections and articles explaining how the work is done came up.

We extended that pattern with our own research into 25 Dutch brands in eight sectors: 1,200 answers from ChatGPT and Gemini. The pattern held. In 73% of the answers with sources there was no about page. About pages made up a quarter of the evidence for who a brand is, the brand’s own offer and price pages almost a third of the evidence for value, and review sites and the brand’s own complaints and service pages most of the evidence for trust. The outside sources differ per sector: the table under step 3 lists the starting points we see most.

The source grader therefore links every row to the claims it supports:

The True Root stays the anchor for identity, and links to the evidence for every other claim: together they form your evidence architecture. The question shifts from which page to optimise to what evidence an AI needs for each claim you want it to make, where that evidence lives, and whether it tells the same story everywhere. Some of it sits on subdomains run by support, HR or marketing teams; AI treats those as part of the brand all the same.

The research shows which pages are cited, not that adding a page causes more citations, and it covers publishers only. Treat the claim map as something to audit and test.

Every row is one of two types, and the two types ask for a different commitment.

Some sources appear twice, once for each type. Replying to reviews is hygiene; running a programme that brings in new reviews every month is distinctiveness. The same goes for forums and the press: correcting errors is hygiene, building a presence is distinctiveness.

The advice: do all the hygiene, then choose two or three distinctiveness rows, give each an owner and a fixed cadence, and do those very well. Two bets kept up for a year do more than ten that stop after the first round.

Choose a bet when two things are true. The brand has a unique asset for it, such as its own data, access or expertise. And the brand can keep up the cadence for at least twelve months.

The order of the work matters more than any single row. The source grader follows the cascade from your own source to the sources that copy it.

1. The root first. Your entity home, your canonical description and your structured data come first, because every other source copies from them. A register, a profile or an AI answer is only as right as the source it was copied from.
2. Then the copies of the root. Registers, Wikidata, business profiles and partner pages follow, with the same sentences and the same facts. The same fact in many independent places is what makes it stick.
3. Then your bets. The distinctiveness rows you chose come after the brand is clearly identified. AI systems split pages into fragments, and a unique figure on a page where the brand is unclear can end up credited to someone else.
4. Then monitoring. Reviews, forums and the press come last, because a correction is strongest when it can point to a source that is already right.

Prioritising happens in three layers: showstoppers first, then four waves in a fixed order, and within each wave a RICE score per action.

Showstoppers first

A showstopper is an error that already produces a wrong or empty picture of your brand. It goes before everything else.

• Identity contradiction: your own site or structured data states a fact that contradicts the company register.
• Entity confusion: AI answers or search results describe another company, an old name or a namesake as your brand.
• Invisible facts: key facts exist only via JavaScript or as images, so AI reads them as missing or as zero.
• Harmful claim: a wrong or damaging claim sits in a source that AI cites often.
• Blocked access: AI crawlers cannot fetch your entity home or key pages.

A block can be deliberate, and from outside a deliberate choice and a mistake look the same. Ask whoever runs the site before treating a block as a showstopper. When you find more than five showstoppers, the root itself needs rebuilding: treat all of wave 1 as urgent.

Four waves

RICE within each wave

Within a wave, rank the actions with RICE, a prioritisation method developed at Intercom: reach times impact times confidence, divided by effort. For distinctiveness, RICE also helps you choose: rank the candidate rows, then test the top ones on the two questions above.

• Reach (1 to 10): How many relevant AI answers and search results the source appears in. Measure it where you can: count how often the source shows up in the answers and results for your key prompts and queries.
• Impact (1 to 10): How much the picture of your brand changes when this source is right.
• Confidence (0.5, 0.8 or 1): 1 when you have seen this source cited for your own brand, 0.8 when other cases or studies show the effect, 0.5 when it is an assumption.
• Effort (1, 3, 5 or 10): The work needed to get it right.

As a rule of thumb, actions in the circle of control take little effort and have moderate impact, and actions in the circle of concern have the highest impact and take the most effort. Three examples:

• Correcting a wrong head office in your own structured data: 8 × 6 × 1 / 1 = 48.
• Correcting your profile on a comparison site: 6 × 7 × 0.8 / 3 = 11.2.
• Requesting a correction on Wikipedia through the talk page: 7 × 9 × 0.8 / 10 = 5.0.

RICE ranks actions within a wave only. A high score never moves an action into an earlier wave.

1. Write down your reference facts. One sentence that says what your brand is, a short and a long description, and the fixed facts: legal name, address, founding date and category (row C02). Every row is checked against this set.
2. Write down your claims. Three to five things you want AI to say about you, at least one each for identity, trust, value and expertise.
3. Measure a baseline. Write a few prompts per claim, each in a short version as a reader would type it and a longer, comparative one. Run every prompt at least twice in at least two AI assistants, with web search on, and record which sources each answer cites (row X01). Short prompts often return no sources at all; the long ones show which sources a model actually uses.
4. Sort the cited sources. A source that comes back in both runs and in more than one assistant is a core source. A source that shows up once goes on a watch list. Core sources get an action; the watch list gets checked again in step 2.

Go through every row in the hygiene list below, compare what you find with the action, and give the row one mark. Start with the rows that cover your core sources. Wikipedia and the main review platform, such as Trustpilot, come back in every sector, so check those in any case (X08, X05). Then work in the order set out above: showstoppers first, then the waves, and RICE within each wave.

Ask first exists because a deliberate choice and a mistake look the same from outside. An insurer may hide old policy documents on purpose, because only the current version on the website should count.

Hygiene is set right once and then kept right. Give the hygiene list one owner and repeat two things every quarter.

1. Run the baseline prompts again. Same prompts, same assistants, at least two runs. Compare the cited sources with the previous round. A source from the watch list that now comes back in two runs becomes a core source and gets its row checked. A core source that disappears needs no action.
2. Recheck the hygiene rows that change by themselves. Reviews, forums, Wikipedia, comparison sites and registers change without you. Recheck those rows and mark them again.

The hygiene score is the share of applicable hygiene rows marked OK, overall and per circle. The target is 100. Rows marked ask first count once they are answered, and rows marked not applicable do not count. Report how many rows you could check next to the score, so a high score on a partial check is not mistaken for a complete one.

Start choosing when most of the hygiene list is OK. A unique claim on a page where the brand is unclear can end up credited to someone else.

1. Find your weakest claim. The claim whose prompts cite your own website least in the baseline.
2. List the candidates. The rows in the list below that support that claim, plus the rows that cover a core source from your baseline.
3. Keep two or three. Keep a candidate only when both are true: you have a unique asset for it, such as your own data, access or expertise, and you can keep up the cadence for at least twelve months. Rank what is left with RICE.
4. Commit. Give each bet an owner, a cadence and the prompts it should move. Mark every other row as not chosen and review them when a bet is finished.

The core sources differ per sector. These are the starting points we see most:

For each bet, keep five things up to date: the owner, the cadence, the date of the last publication, whether the schedule was met, and the prompts the bet should move. Check those prompts in every monitoring round. A bet that misses its schedule twice needs a new owner or a different choice.

• Source type: The kind of source, with example platforms and the claims it supports: identity, trust, value or expertise.
• Action: What to do, and where.
• Further reading: Articles, studies and guidelines that go deeper, with a note where one applies.

Each row has an ID: C for the circle of control, I for the circle of influence and X for the circle of concern. Link to a row directly with its ID, for example #c01.

The source grader anchors the sources that AI draws from, which is the part a brand can work on. What an AI model then says is measured afterwards and never promised. The ranking in this method is a first calibration, refined with every engagement.