GEO content tiers: facts, opinions, promises — which one does AI cite first?

Bottom line first: When AI writes a summary, it grabs verifiable specifics, not adjectives. Think of your content in three tiers: the fact tier carries scope, conditions and numbers; the opinion tier offers judgments and trade-offs; the promise tier is guarantees, leadership claims and 100%. The first two can be excerpted; the promise tier almost never gets cited. The smallest fix is to add a condition and a number to every adjective.

1. What AI excerpts are “verifiable sentences,” not adjectives

Paste the first paragraph of your homepage into an AI and ask it to summarize “what this company does.” If that paragraph says “a professional, leading, trusted partner,” all you get back is a vague paraphrase, because there is no extractable fact in the sentence.

If instead the paragraph says “we make aluminum extrusion dies, focused on 6xxx and 7xxx series, with annual capacity of about 1,200 sets and a minimum order of 50 sets,” the model can lift those fields straight into its answer.This isn’t model bias — it’s sentence structure— the first has nothing that can be matched against other sources; the second does.

Sentence type	Typical wording	What the model can extract	How it usually gets handled
Verifiable statement	Tolerance holds at ±0.02 mm, sampling per AQL 1.0	Values, units, standard numbers	Can be quoted directly and cross-checked against third-party sources
Judgment and advice	Under 500 pieces, go with a rapid tool; don’t cut a steel die	Stance, conditions of use	Can be quoted, usually attributed as “one view holds”
Promises and self-assessment	Industry-leading, 100% guaranteed, serving customers worldwide	None	Mostly ignored, or flagged as marketing language
Completely empty	We put our heart into every product we make	None	Not cited at all

Using this table is simple:After you write a paragraph, ask yourself: “Could someone else take this sentence and verify it?” If it can be verified, it’s the fact tier. If it can’t be verified but reveals your judgment, it’s the opinion tier. If it can neither be verified nor shows judgment — just attitude — it’s the promise tier.

2. How to tell the three tiers apart: markers and typical sentence patterns

Tiers aren’t about how strong the tone is, but whether the sentence has a “verifiable anchor.” There are four kinds of anchors: scope, conditions, numbers, sources.

Tier	Marker	Typical sentence pattern	Where it appears
Fact layer	Contains at least two of scope, conditions, numbers, sources	Under what conditions, within what range, against what standard	Spec tables, process descriptions, FAQ, delivery notes
Opinion tier	Makes a clear call and states when it applies	We recommend A because of B; choose C when…	Buying guides, comparison posts, how-to content
Claims	Only self-praise or absolutes, nothing verifiable	leading, guaranteed, best, 100%, always	Homepage banners, sales pitches, proposal covers

Here’s a practical test:Cross out every adjective in the paragraph and see what’s left. If what’s left is specific conditions and numbers, it’s a fact. If it’s a clear call, it’s an opinion. If nothing’s left, it’s a claim.

One boundary here:You sort sentences, not whole articles. All three types usually appear in the same article, and that’s fine. What matters is putting each one where it belongs: facts as the backbone, opinions as the calls, claims cut to zero or given limits. Writing everything as bare facts isn’t realistic and makes for painful reading.

3. How quotable each type is: why claims almost never get cited

The three types get cited at very different rates. It’s not that models “hate adjectives” — it’s that claims can’t be matched against anything else.

Tier	Chance of being cited	Why	How to rewrite
Fact layer	High	The sentence carries its own context, stands alone, and can be cross-checked against other sources	Keep it, add scope and source
Opinion tier	Medium	Can be cited as “one approach,” but needs supporting facts	Attach a fact to every opinion
Claims	Extremely low	Unverifiable, indistinguishable from the same words from any other supplier	Break into condition + metric + range
Absolutes	Close to zero	May trigger compliance risk; models tend to avoid them	Delete entirely or add scope of use

Claims carry a hidden cost:They don’t lose points on their own, but they drag down the credibility of every other sentence on the page. When a page says both “±0.02 mm tolerance” and “industry leader,” the model treats the whole page as marketing material during cross-checking, and even solid facts get discounted. That’s why the same fact gets cited more often on a clean page than on a promotional one.

4. Eight rewrites: from claims to facts

These are the eight sentences we rewrite most often on client sites. The original is on the left, the rewrite on the right. The difference isn’t style — it’s the conditions, ranges and scope that got added.

Original claim	Rewritten as fact	What was added
We are an industry-leading precision machining supplier	12 years in CNC precision machining, mainly aluminum and stainless steel parts, tolerances held at ±0.02 mm, about 400,000 units shipped per year	Years, materials, tolerance range, capacity
Fast lead times you can trust	Prototyping 5–7 days for standard parts, production 15–20 days including finishing; add about 5 days during the September–November peak	Stage-by-stage timelines, peak-season variance
Guaranteed quality, 100% inspected	Before shipment, we inspect critical dimensions by AQL 1.0 sampling. Only sealing faces and mating holes get 100% inspection. A CMM report ships with every order.	Sampling plan, full inspection scope, documentation
Our pricing is competitive	At 500 units, unit prices for comparable specs run 8%–12% below typical market quotes. The gap comes mainly from our in-house heat treatment line, which cuts outsourced processing.	Volume threshold, price gap, source
We serve customers worldwide	Our customers are in Germany, Poland, the US, and Vietnam. Europe accounts for about 45% of revenue. We deliver in English and German.	Markets, revenue share, languages
Our engineering team has deep experience	The engineering team has 6 people, 3 of them with 8+ years on the same processes. They handle DFM reviews and first-article approval, with a 2-business-day review turnaround.	Team size, experience, responsibilities, turnaround
Our products meet international standards	Dimensional tolerances follow ISO 2768-m and customer drawings. The ISO 9001 certificate number and expiry date are on our certifications page for verification.	Standard number, scope, where to verify
Worry-free after-sales support, always responsive	We reply to emails within 4 hours on business days, 8:30–17:30 (GMT+8). During the warranty period, we replace non-human-caused failures free of charge, and we cover shipping both ways.	Service hours, response time, what’s covered

The most common mistake when rewriting is mistaking “more specific” for “write more”. Rewriting isn’t about adding adjectives; it’s about adding constraints. If you rewrite “fast lead times” as “very fast lead times, usually delivers quickly,” it’s still worthless.

The second common mistake is giving ranges that are too wide. Writing “tolerances from ±0.001 to ±0.5 mm” looks like broad coverage but actually says nothing. Narrow the range to what you can reliably hold. For anything you can’t, write “requires evaluation.”

5. Four sentence templates for the facts layer

Fact sentences don’t need to be fancy. Just follow a template. These four cover 80% of B2B website scenarios.

Template	Structure	Where it applies	Example
Condition + metric + range	Under what conditions, what range of values	Spec pages, capability pages	For lengths under 600 mm, flatness can be held within 0.05 mm
Method + turnaround + result	What method, how long, what result	Process pages, workflow pages	First articles are fully dimensioned on a CMM, with a report in 2 business days
Standard + test conditions + conclusion	Which standard, tested under what conditions	Certification pages, quality pages	Tested to GB/T at room temperature: 48 hours of salt spray with no red rust
Comparison + difference + when to choose	Compared to what, by how much, when to pick which	Selection guides, comparison articles	Compared to die casting, CNC costs about 20% less per unit at low volumes and suits runs under 500 units

Of the four templates,the first and fourth are worth filling in first. The first answers “what can you do,” and the fourth answers “when should a buyer come to you.” Both types of sentences get used by procurement teams and AI models alike.

Two things to watch when writing: spell out units in full (mm, kg, days, hours) — don’t drop the unit and leave just a number; and keep definitions consistent. In the same article, “lead time” can’t include surface finishing in one place and exclude it in another. Factual statements with inconsistent definitions will be flagged as unreliable during cross-validation.

6. How to use the opinion layer: attach it to facts

The opinion layer isn’t meant to be deleted. Quite the opposite:a page with no opinions reads like a spec sheet, and conversion is driven by judgment. The problem is that opinions need to be attached to facts, not left floating.

How it’s written	Example	Issue	How to fix it
Standalone opinion	For small batches, we recommend rapid tooling	No premise given — the reader doesn’t know what counts as a small batch	Add the condition: under 300 pieces
Opinion + fact	For orders under 300 pieces, we recommend rapid tooling; the tooling cost is about one-third that of steel tooling	Citable, with support for the judgment	Keep as is
Fact + opinion + boundary	For orders under 300 pieces, we recommend rapid tooling; the tooling cost is about one-third that of steel tooling. But it doesn’t apply when tolerances must be within ±0.02 mm	Complete, ready to quote directly	Use as the target pattern

The target structure I recommend is “one fact → one judgment → one boundary“. Three sentences, under a hundred words — the model can lift the conclusion directly, and readers can immediately tell whether it applies to them.

There’s one exception to this rule:in tender documents, formal quotes and contracts, the opinion layer has no place. That kind of text needs only facts and clauses; adding “we recommend” actually comes across as unprofessional. Layering is a method for content pages, not a universal standard for all documents.

7. Layering self-check list and boundaries

After rewriting a page, run through the eight items below. Half an hour is enough.

Checklist	Qualifying threshold	Common signs it falls short
Share of factual sentences	Body text ≥60%	The whole piece is philosophy and vision
Does each paragraph have an anchor?	At least 1 verifiable statement per paragraph	Three paragraphs in a row with no numbers
Are promise-style sentences at zero?	Leading, best, 100%, guaranteed appear 0 times	Only synonyms were swapped; the meaning didn’t change
Do absolute statements have a scope?	Every one carries its conditions of use	Says “all” without saying what all includes
Are number definitions consistent?	One definition per metric throughout	Lead time includes outsourcing in one place and excludes it in another
Do opinion sentences have boundaries?	Every judgment states when it doesn’t apply	Only recommendations, no exceptions
Can the conclusion section stand on its own?	Reads clearly when pulled out on its own	Only makes sense by referring back to earlier text
Are credentials and numbers verifiable?	Certificate name + number + validity period + lookup link	Only says “certified”

Here’s the priority order for going back and fixing old articles:

Priority	What to fix	Page type	Expected impact
P0	Promise sentences on the homepage and about page	Homepage, about page	Affects entity recognition and overall credibility
P1	Fact gaps on product and capability pages	Product pages, process pages	Directly affects the odds of being quoted
P2	Opinion boundaries on selection and comparison pages	Guide pages, comparison pages	Improve citations in the decision stage
P3	Align FAQ and delivery notes on wording	Support pages	Reduce contradictory answers

This tiering doesn’t apply in three cases. First, bidding documents and contracts, where you must use promissory language — the rule there is to add qualifiers, not delete them. Second, scenarios governed by advertising law, where the goal shifts from “easy to cite” to “compliant,” and the two don’t fully overlap. Third, sites still early-stage with few facts to work with — if you don’t yet have numbers for your production lines, testing capabilities, or customer mix, don’t rewrite the copy first; go collect the data. Without facts to draw on, any tiering technique is just a different way of saying nothing. For related reading, see Trust pages and AI credibility and read that one alongside this. Author Cui Lishu In delivery, this step usually sits in pre-launch QA and takes about 4–6 hours to run through the whole site. If you need a tiering audit against your content list, you can Contact Us.

FAQ

What share of sentences should be factual?

For content pages (process, selection, FAQ), aim for at least 60% of the body to be verifiable statements; the homepage and about page can go down to 40%, since those two need to state positioning and stance. But there’s one hard floor: on any page,the count of absolute claims must be zero. The test is simple: cross out every adjective in the body, and if what’s left is still a complete sentence, you pass.

What if there really are no specific numbers?

First work out whether it’s “none” or “not counted yet.” Capacity, lead times, testing standards, customer distribution — nearly every factory has this data, it’s just never been organized. Two days is usually enough to pull the main metrics together. If a particular item genuinely can’t be disclosed, fall back to ranges and conditions, e.g. “Batch size and lead time are assessed per drawing; standard reply within 2 business days.” That’s still a verifiable, promissory fact, and far more useful than an adjective.

Do industry-standard statements count as promissory?

It depends on whether they’re tied to you. “Castings require controlled cooling rates” is industry common knowledge — a neutral statement that can be cited. “We use advanced cooling control technology” doesn’t say what you actually do, so it’s promissory. The difference is whether the subject of the sentence describes you. Describing industry norms is safe to write; describing your own capabilities requires a number or a standard code to back it up.

What if a customer explicitly asks for “guaranteed” or “100%”?

Handle it case by case. Contracts and letters of commitment follow whatever the client asks for, since those are legal documents and not subject to this tiering. On the website and in marketing materials, the approach is to replace absolute claims with verifiable alternatives: change “100% inspected” to “critical dimensions sampled at AQL 1.0, sealing surfaces 100% inspected”; change “guaranteed 15-day delivery” to “15–20 days standard, about 5 days longer in peak season, delays handled per contract terms.” This keeps your commitments credible and keeps them quotable.

Will AI quote opinion statements?

Yes, and often, but it quotes them differently. Factual statements usually get lifted straight into the answer, while opinion statements are more often paraphrased as “one approach,” for example “some sources suggest using a rapid mold for runs under 300 units.” That means opinion statements show up in answers more easily, but it’s harder to get your brand attached to them. So the value of an opinion statement lies mainly in the decision stage, and it’s best placed right after a factual statement, so the model picks up your conditions and data along with the excerpt.

Should old articles be revised?

Yes, but not all of them. Start with the 10–15 pages that get quoted and visited the most, meaning the homepage, main product pages, and core process pages. Don’t rewrite the whole page; just fix three types of sentences: absolute claims, promises with no conditions, and numbers that don’t match across pages. After one round, track only one metric: whether the same batch of questions gets more accurate factual descriptions of you in next month’s answers. For a closer look, pair this with Correcting brand misinformation the method in that article.

Final thoughts

Content tiering, at bottom, answers one question:Can someone else verify this sentence? A verifiable sentence automatically earns the right to be quoted; an unverifiable one, no matter how well written, just wears out the reader’s patience. AI simply magnifies that difference, because it has no patience and no sympathy.

Here’s the order I’d suggest: run a one-page pilot first, strip out the promise sentences and add conditions and numbers; then a week later, see whether the answers to the same batch of questions describe your facts more accurately.The payoff here isn’t in the traffic curve. It’s whether the few sentences people repeat about you are correct.

One line worth remembering: verifiable statements get repeated, adjectives don’t.