Reading a Google Patent Without Fooling Yourself: Claims First

Google Patents puts the description above the claims. You land on the page, scroll past the abstract, and the next several thousand words are readable prose full of quotable sentences. The claims sit underneath all of it, in numbered legalese most people never reach.

That layout explains most of what is wrong with patent commentary in this industry. The description is the part no examiner ever tested; the claims are the only text that had to survive one. Almost every patent take you read this year quoted the wrong half.

What a patent is evidence of

A granted patent is evidence that an applicant wrote a description and a set of claims, and that an examiner allowed those claims over the prior art they found. The USPTO is explicit about what the grant confers: the right to exclude others from making, using, offering for sale, selling or importing the invention. No obligation to build anything, and no evidence that anything was built. One inference survives: somebody at the assignee thought this was worth attorney time and a multi-year prosecution. Intent, not systems.

Four inferences it cannot support, in the order people make them:

That the mechanism ships. Large companies file defensively, file to block competitors, and file around ideas they abandoned. The grant records none of that.

That it ships now. Three dates matter and they can be a decade apart. US9418104B1 has a priority date of 2009-08-31, a filing date of 2012-09-14 and a publication date of 2016-08-16. Read in 2026, its idea is seventeen years old.

That it ships as written. Claims are drafted for scope, not accuracy, and the description enumerates alternatives nobody built because that preserves room for continuations. Breadth is a drafting goal; you are reading a drafting artifact.

That the named quantity is a ranking factor with a weight. A claim describes a procedure and a data flow, not a magnitude. If your take contains a number that is not in the document, you invented it.

The front page, field by field

Before reading a word of the text, read the metadata block. Four fields move the weight of the document.

Assignee, current and original: Google Patents prints both, with a disclaimer that “the listed assignees may be inaccurate.” A patent that arrived through an acquisition is much weaker evidence about Search than one prosecuted in-house; Google Inc. to Google LLC is only the 2017 rename.

Priority, filing and publication dates. The age of the idea, when the application entered the queue, and when you were allowed to see it. None of them is a deployment date.

Legal status, and the kind code in the number. A1 is a pre-grant publication of an application; B1 and B2 are granted patents. A large share of “Google patented X” posts are about A1 documents that may never have been allowed. Google prints the status field itself with a warning worth quoting in full: “The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.” Four values you will meet:

Maintenance fees fall due at 3.5, 7.5 and 11.5 years after grant, with grace periods after each (USPTO). At a company with a patent department, a lapse is not an oversight but a line item somebody declined to renew, and the Legal Events table at the bottom of the page is the receipt.

Family and continuations, the strongest front-page signal and the one nobody looks at. US7716225B1 sits in a family of four applications sharing a 2004-06-17 priority date, filed in 2004, 2010, 2012 and 2016. Twelve years of refiling on one idea is a company that kept paying attorneys. Compare a single grant with no children: one filing, one decision, no follow-through.

Claims versus description

The statute separates the two texts by function. 35 U.S.C. 112(a) requires a description good enough for a skilled person to build from. Subsection (b) requires that the specification “conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.”

Only the second text was negotiated. Examination is an argument about claim scope: the examiner rejects, the attorney narrows, what survives is the claim set. The description is filed once and never trimmed, because trimming would surrender support for continuations. It is over-inclusive on purpose, and you can measure that in ten seconds with stdlib:

import re, html, urllib.request

URL = "https://patents.google.com/patent/US8682892B1/en"
req = urllib.request.Request(URL, headers={"User-Agent": "patent-reader/1.0"})
page = urllib.request.urlopen(req).read().decode("utf-8")

def section(name):
    m = re.search(r'<section itemprop="%s".*?</section>' % name, page, re.S)
    return re.sub(r"\s+", " ", html.unescape(re.sub(r"<[^>]+>", " ", m.group(0))))

PATTERNS = [r"\bmay\b", r"\bcan be\b", r"one or more", r"embodiment", r"implementation"]
for name in ("description", "claims"):
    text = section(name)
    print(name, len(text.split()),
          [len(re.findall(p, text, re.I)) for p in PATTERNS])

Against US8682892B1, the patent this post works through:

Eighty-eight hedges in the half that was never examined, zero in the half that was. The ten remaining “one or more” in the claims are structural (“one or more computers”), not optional features. Run it on whatever patent is on your timeline and check the shape yourself.

That is the observation. My reading of it: the description is a wish list written by someone paid to keep options open, and quoting it is closer to quoting a pitch deck than a specification.

From claim language to a falsifiable prediction

The translation is mechanical. Four steps.

1. Take the independent claim, alone. It stands without reference to any other, and everything else narrows it. Number its steps: each is a separate thing that must be true.

2. Replace each term with the narrowest meaning a dependent claim gives it. Dependent claims are definitions that survived examination; the description’s were never contested. Search the claim set before you search the description.

3. Write the prediction as a population, a direction and a comparison. “If this is running, then across population X, holding Y constant, Z should differ in direction D.” If you cannot name the population, you do not have a prediction, you have a mood.

4. State what you would observe if it were false. If nothing would count against it, stop.

Some vocabulary tells you in advance that step 3 will fail:

• “may”, “can be”, “in one embodiment”, “in some implementations” introduce alternatives rather than commitments, and per the table above they are absent from claims. If the sentence you want to quote contains one, you are quoting the description. Stop there.
• “one or more of” is a disjunction. Any subset satisfies it, including the one-element subset, so no observation contradicts it.
• “based on” says two quantities feed a third, and nothing about how. Any monotone function qualifies, so it supports a prediction about sign, never about magnitude.
• “wherein X is a ratio of A to B” is testable, and exactly the language that lives in dependent claims. That is where to go looking.

Designing the test

Take each term in the claim and write down the closest thing you can measure from outside, and the error you accept. If a term has no proxy, the test dies there.

Start with the observable proxies. Rank is observable. Referring domains are semi-observable through a third-party link index, at a coverage you cannot quantify. Query volumes are not: what tools sell as search volume is a model output. Anything internal to scoring is not observable at all, which rules out most patents before you start.

Sample size, computed rather than guessed. The unit is usually the (query, URL) pair, not the site, and correlations there are small. For a rank correlation test the Fisher transform gives the requirement directly:

import math
za, zb = 1.959963985, 0.8416212336   # two-sided alpha 0.05, power 0.80
for rho in (0.15, 0.20, 0.25, 0.30):
    print(rho, round(((za + zb) / math.atanh(rho)) ** 2 + 3, 1))

Detecting a correlation of 0.20 needs 194 pairs; 0.15 needs 347. If your design collects thirty, you built an instrument that cannot see the effect, and its null result means nothing.

Then the control. The best control is not a random sample. It is a second prediction, from the same claim set, of where the effect should be absent. Patents are generous here: they are full of exceptions and thresholds, and each one marks a place your effect must vanish. If it shows up everywhere, you are measuring a confound.

And the honest verdict: “no test exists” is a legitimate and common outcome. Most patents describe mechanisms whose inputs are internal, whose outputs are entangled with a hundred other systems, and whose predicted effect is smaller than the daily variance of the SERP.

Worked example: US8682892B1, “Ranking search results”

US8682892B1. Inventors Navneet Panda and Vladimir Ofitserov. Original and current assignee Google LLC. Priority and filing date 2012-09-28, granted 2014-03-25, application US13/631,020.

Legal status: Expired - Fee Related, expires 2032-10-06. The Legal Events table shows the 4th-year maintenance fee paid 2017-09-25, a reminder mailed 2021-11-15, and a lapse for failure to pay effective 2022-03-25. Google paid once and then declined, ten years before the term would have run out. Family: one application, no continuations, US only. Forward citations: seven, none of them Google’s.

That is the front page of a document the industry still cites as “the Panda patent”. The word “Panda” appears nowhere in its abstract, description or claims; it is the first-named inventor’s surname. Google’s guide to ranking systems lists a retired Panda system “announced in 2011 and given the nickname of the ‘Panda’”, which “evolved and became part of our core ranking systems in 2015”, a nickname older than this filing. Both are facts; the bridge between them is an assumption about a surname.

Claim 1, verbatim

1. A method performed by one or more computers, the method comprising: determining, for each of a plurality of groups of resources, a respective count of independent incoming links to resources in the group; determining, for each of the plurality of groups of resources, a respective count of reference queries; determining, for each of the plurality of groups of resources, a respective group-specific modification factor, wherein the group-specific modification factor for each group is based on the count of independent links and the count of reference queries for the group; and associating, with each of the plurality of groups of resources, the respective group-specific modification factor for the group, wherein the respective group-specific modification for the group modifies initial scores generated for resources in the group in response to received search queries.

Four steps: count links, count queries, combine them into one factor per group, apply it to scores. Note what claim 1 does not say. Not how the two counts combine: “based on” covers every function you can imagine. Not whether the factor demotes or promotes. Not what a group, a reference query or independence is.

The dependent claims that define the terms

Claim 5 identifies the group “based on an Internet address”, so a group is a host, not a topic cluster. Claim 16 makes a link independent when source and target are “included in different groups”. Claim 19 defines a reference query as “a previously submitted search query that has been categorized as referring to a resource in the particular group”. Every definition you need is in the claims.

Claim 21 is the one worth the whole document:

21. The method of claim 1, wherein determining a respective group-specific modification factor for a particular group of resources comprises: determining an initial modification factor for the particular group of resources, wherein the initial modification factor is a ratio of a number of independent links counted for the particular group to the number of reference queries counted for the particular group.

A ratio: independent links divided by reference queries, per host. That is the sentence people paraphrase as “links to brand mentions”, and the paraphrase is wrong in a load-bearing way. The denominator is not mentions on other people’s pages. It is queries typed into the search engine.

Claim 8 exempts navigational queries: the factor “does not modify the initial score”. Claim 10 does the same below a first threshold; claims 11 and 12 give formulas for the band above it. The mechanism is inert at both ends.

Meanwhile, the sentence everyone quotes from this patent, “Search results identifying low-quality resources can be demoted in a presentation order of search results returned in response to a user’s query”, appears in the description, under a lead-in reading “The subject matter described in this specification can be implemented in particular embodiments so as to realize one or more of the following advantages.” Three hedges in one sentence, introducing a list of hoped-for advantages rather than a limitation. The word “quality” appears zero times in the claims.

The prediction

From claims 1, 5, 8, 10 and 21, stated so it could be wrong:

Across non-navigational queries, among co-ranking hosts binned by reference-query volume, position should be associated with the host-level ratio of independent referring domains to brand-name query volume. And that association should disappear on navigational queries and among results at the top of the score distribution.

The second half is the useful half: a site-wide multiplier with a floor threshold and a navigational exemption predicts a specific non-linearity, and confounds like topical relevance do not produce that shape.

The test that would settle it, and why it does not exist

Unit of analysis: the (query, host) pair, at least 194 per bin by the arithmetic above. Independent variable: independent referring domains, deduplicated to one per source host and excluding sources sharing ownership with the target, which is the same edge-list hygiene as computing PageRank over your own crawl graph, applied to an external index. Dependent variable: position. Control: the same measurement on navigational queries, where claim 8 predicts nothing.

It fails at the denominator. Reference queries are queries submitted to Google, and nobody outside Google observes them. The nearest proxies (third-party brand volume estimates fitted to clickstream panels, or Google Trends, which is relative and normalized) correlate with referring domains and with everything else that predicts rank, so the ratio you compute is not the ratio in the claim and a positive result is indistinguishable from “bigger brands rank better”. It fails again at the initial score: claim 1 modifies a number you never see, and claim 10’s threshold is defined on that same invisible quantity, so the control bins cannot be constructed.

Verdict: no test exists from outside. I am not reporting a null result, because I have not run anything; the design above is what would have to exist, and it does not. The correct action after reading this patent is to change nothing. That outcome is available for most patents, and publishing it is the part nobody does. It generalizes: fingerprint the SERP you can actually observe before you theorize about the scoring you cannot.

The other evidence classes, and how each gets overread

Ordered by the weight each deserves, patents included:

Official documentation first. Search Central is the only class you can check yourself, because most of it describes observable behaviour. Diff it over time: the edits are informative, the omissions are not admissions. It still beats folklore, which is why the two systems behind one indexing label can be told apart from the docs alone.

The rater guidelines. Their canonical URL is guidelines.raterhub.com/searchqualityevaluatorguidelines.pdf, which redirects to a copy on static.googleusercontent.com whose rel="canonical" header points back. That file’s Last-Modified, 2025-09-11 as I write, beats any blog post claiming to track versions. Google’s own description, from the Search Central announcement: the guidelines “are what are used by our search raters to help evaluate the performance of our various search ranking systems, and their ratings don’t directly influence ranking.” A definition of the target, not a description of the machinery aiming at it.

Court exhibits are produced under legal compulsion and tested by cross-examination; the DOJ publishes them in full. The overread is temporal: a deck from 2017 describes 2017, and it was one team persuading another.

Leaked internal material. The 2024 case is checkable: the documentation surfaced through automated regeneration commits to a public Google client-library repository in March 2024. Evidence of field names, types and comments in one schema at one moment. Not evidence of which fields are populated, which are read at serving time, or what they weigh. A schema is a vocabulary, not a program. One line of that schema, siteAuthority, graded claim by claim shows how little a field name carries once you refuse to read the commentary back into it. The same standard applies to claims about AI answer surfaces, where reasoning from vocabulary is more tempting because the measurement is harder.

The checklist

Before you post about a patent:

1. Granted patent or application? Check the kind code in the number.
2. Who is the assignee, and did they file it or acquire it?
3. What is the priority date, and how old is the idea today?
4. What is the legal status, and what do the Legal Events say happened?
5. Is there a family? Continuations still being prosecuted, or one filing and silence?
6. Are you quoting a claim or the description? Search your quote for “may”, “can be”, “one or more of” and “embodiment”.
7. If it is a claim, is it independent, and have you read the dependent claims that define its terms?
8. Can you state the mechanism as a population, a direction and a comparison?
9. Is every term in that statement observable from outside, at what proxy and what error?
10. What sample size does the predicted effect require, and can you collect it?
11. What would you observe if the claim were false? If nothing, say “no test exists” and change nothing.

Most patents die at check 6. Of the survivors, most die at check 9. That is not a reason to stop reading patents. They are one of the few windows into what a search company thought worth protecting, and the drafting tells you which problems they found hard. It is a reason to stop treating them as descriptions of a system you are optimizing against. The cost of that is not embarrassment on a timeline. It is a quarter of somebody’s roadmap spent chasing a ratio whose denominator nobody outside Google has ever seen.