UNmiss Blog

How RTINGS Outranks Forbes and CNET With a Machine-Built Page Layer

How RTINGS turns first-party product testing into useful comparison pages that rank above major publishers, plus the limits of the model.

RTINGS beats Forbes and CNET on Google for product questions people ask just before they buy. It does not win by trying to publish more opinions than they do.

Picture someone choosing a TV for a bright living room. They do not want ten tabs and an evening of research. RTINGS puts the test results and the answer to "which one is better?" in one place, then repeats that useful job across products and comparisons.

That simple idea is the story: start with evidence shoppers need, then make comparisons easy to use. The measurements later in this case study show the scale; they are not the reason a reader should care first.

Then we asked AI assistants twelve buyer questions in the same categories and found the one place the playbook stops working. That answer is why this case matters now rather than in 2023.

Below: what actually produces the result, what you can copy this week, what you cannot copy at all, and the six-step version you can run on a competitor this afternoon.

Results at a glance

#1
on Google for "best tv for bright room", above Reddit, Forbes and CNET
722,598
backlinks from 35,625 domains, earned without outreach
0 of 6
AI answers naming them, when the model did not search
The short version

RTINGS earns strong search visibility because it turns original lab tests into comparisons shoppers can use. Its link profile and comparison layer support that work. Our AI check showed a separate problem: being easy to find in search does not guarantee that a model remembers to name you without searching.

One source note before the findings. Our own Anatolii Ulitovskyi published an analysis of this site on LinkedIn in December 2025. He fact-checks this blog, and a handful of figures below are credited to him and were not re-verified by us. Every other figure we measured ourselves on 13 August 2026, with free tools, and each one is visible in a screenshot you can open at full size.

The RTINGS model

RTINGS buys consumer electronics at retail, tests them on the same instruments every time, and publishes the measurements. TVs, headphones, monitors, mice, keyboards, printers, blenders, running shoes.

The important part is that they publish the raw numbers rather than a summary. Once every TV in the database has a measured contrast ratio, response time and peak brightness, a page comparing any two of them can be rendered without a writer. The content is not spun — it is data that already existed, laid out.

That is what the sitemaps look like when you count them.

Why each review creates more useful comparisons

This is the central reason the strategy works, and it is arithmetic rather than SEO. Reviewing the thirtieth TV in a category does not add one page. It adds twenty-nine — one comparison against every TV already in the set.

The evidence is the ratio above: 6,628 editorial URLs supporting 22,487 generated ones. The editorial pages are not the product. They are the input to the product.

That inversion is what most teams copying programmatic SEO get backwards. They start with the templates and go looking for data to fill them. RTINGS started with an instrument, and the templates are a consequence.

Your version: before you plan a single template, ask what your business measures first-hand that nobody else has — real prices you paid, real delivery windows, real failure rates. If the honest answer is nothing, the rest of this playbook will not work, and the sections below say so plainly.

Turn a comparison into a page

The unit of production is not an article. It is a rendering. Here is one of the 22,487, pulled from the comparison sitemap.

An RTINGS side-by-side comparison page for two OLED televisions, showing the two products next to each other with their measured test scores laid out in matching rows for direct comparison.
What to notice: there is no prose. Two products, the same measured fields, in the same order, with the same units — so the reader's eye runs down a column and stops where the numbers diverge.

It works because it removes labour the reader would otherwise do. Answering "which of these two" by hand means two browser tabs and twenty minutes of scrolling. The page costs RTINGS nothing beyond tests they had already run.

Your version: pick the comparison your customers make in a spreadsheet before they buy, and render it. If you cannot name that comparison, you do not have a programmatic layer yet — you have a template.

Let the data model do the linking

At 29,000 URLs, internal linking cannot be a person's job, and at RTINGS it is not. Each generated comparison page links back to both product reviews. Each review links to the test methodology behind its numbers. Each methodology page links out to every product it was used on. The link graph falls out of the data model.

We pointed our free Website Audit at their homepage to see how that holds up from a crawler's side.

UNmiss Website Audit result for rtings.com showing a health score of 90 out of 100, 0 critical issues, 1 warning and 3 notices with 118 checks passed, and group scores of 25 of 25 for on-page SEO, 65 of 67 for technical SEO, 27 of 29 for speed and 2 of 2 for mobile.
What to notice is not the 90/100 — it is technical SEO at 65/67. That is the group that breaks first when a site scales, and theirs is intact. The four flags are an LCP warning, missing social card tags, nofollow internal links and uncompressed CSS: housekeeping, not architecture.

Two of those notices reward a second look. Nofollow internal links on a site this size is usually deliberate — it is how you stop crawl budget draining into filter and sort URLs. The missing social preview tags are a genuine miss that costs them nothing in Google and something in every share.

Your version: when you audit a programmatic build, ignore the headline score and read which group loses points. On-page failures are copy problems you can fix per template. Technical failures at scale mean the architecture is not holding, and no amount of content fixes that.

Write clear titles and useful descriptions

RTINGS writes meta descriptions far past the display limit on purpose, but the pattern does not extend to titles. Anatolii's analysis flagged oversized meta tags generally. We sampled 20 pages — 14 reviews and 6 comparison pages — and the measurement sharpens the claim.

The logic holds up. Google rewrites descriptions for most queries anyway, so a 259-character description is not padding — it is source material for a snippet engine that is going to rewrite you regardless. The longest we found ran to 697 characters and read like the opening of the review, because that is what it was. A 90-character title is just a truncated title.

Your version: if you generate tags at scale, write the description as a genuine paragraph of summary and let it run long. Keep titles inside 60. Do not read "write long" as a rule that applies to both fields.

Target buyer-ready questions at scale

RTINGS does not compete for head terms. It competes for the query typed with a credit card already out, thousands of times over. "Samsung S90D vs LG C4" is not a topic anybody writes a blog post about; it is a question asked in the last five minutes before buying, and there are 22,487 of them.

We checked one of their bread-and-butter terms with the free Keyword Research tool.

UNmiss Keyword Research overview for the term best tv for bright room in the United States, showing difficulty 13 rated Easy, 810 monthly searches, cost per click of 31 cents, competition 1.00 and informational intent, with a 12-month trend chart.
What to notice: difficulty 13 and only 810 searches a month. Individually this term is not worth a content brief. Multiplied across every product, size, room type and use case, it is the whole business.

Notice the intent field too: Info. Someone choosing a television for a sunny living room is four minutes from a purchase, and the label still says informational. Anatolii puts informational keywords at 34.3% of RTINGS' traffic, which is the same observation from the other side: the tag on the keyword and the stage of the buyer are different things.

Grouping is the other half of the job. Ten keywords from four of their categories, dropped into the Keyword Cluster Tool, came back as four clusters with pillar topics attached.

UNmiss Keyword Cluster Tool result grouping ten seed keywords spanning TVs, soundbars, headphones and monitors into four clusters, each with a pillar topic and the keywords assigned to it.
What to notice: four pillars from ten seeds. On a real build, that pillar list is your template list — it decides which page types you create and which pages hang off which hub.

Your version: stop sorting candidate keywords by intent tag and start sorting by what the answer has to contain. If the honest answer is a ranked list of specific products, that page belongs in the programmatic layer. If it is an explanation, it belongs in the small editorial set that feeds it.

Let machines format facts, not invent them

Analysing their top eleven pages, Anatolii put the mix at roughly 25% human-written and 75% AI-generated, with the machine handling technical specification writing against structured data. We did not re-run that classification, so treat the split as his figure from December 2025 rather than ours.

Read carefully, it is not "they publish AI slop". The generated layer is where a machine turns measured values into sentences — contrast ratio, response time, peak brightness in nits — and a machine is better at that than a tired writer, because it never mistypes the number. The human quarter is where judgement lives: what to recommend, what the trade-off is, who should buy the cheaper one.

The caveat is the load-bearing one for this whole article. None of that manufactures the measurements. RTINGS' generated content is trustworthy because the numbers underneath came from an instrument.

Your version: draw the line at the same place. Machines write the layer where the facts are already fixed; people write the layer where someone has to decide something. Generate sentences about data you do not have and you have built 22,487 pages of nothing.

Earn links people actually use

At this scale links stop being a campaign, and the ones that arrive look nothing like the ones you would have bought. We ran rtings.com through our free Backlink Analyzer.

UNmiss Backlink Analyzer report for rtings.com showing 722,598 backlinks, 35,625 referring domains, 21,118 referring IPs, 9,853 referring subnets, 530,785 dofollow against 191,813 nofollow, Domain Trust 86 of 100, 2,972 .edu backlinks and 251 .gov backlinks, a link growth chart falling from 906,485 to 695,846, and a top backlinks table led by discussions.apple.com, en.wikipedia.org and gist.github.com.
What to notice is the referring-page list, not the 722,598. It is led by Apple's support forum, four Wikipedia language editions, GitHub and Adobe's community — and nearly all of them are nofollow.

The best links pass no PageRank. Wikipedia's "Contrast Ratio of 2015 TVs" citation and Apple's discussion threads are there because someone needed a number and RTINGS had it. That is what earned links look like in the wild, and it is the opposite of what a link-building campaign produces.

The long tail is what accumulates them. 36,960 separate pages on the domain have links pointing at them — not the homepage. The programmatic layer is not only ranking, it is collecting citations.

And the profile is contracting. The growth panel runs 906,485 down to 695,846 between 15 May and 13 August 2026. Anatolii cited a mid-2024 peak of over 54,000 referring domains; we measure 35,625 today. Different tools count links differently, so those two numbers are not strictly comparable — but the direction inside our own chart is unambiguous, and it corrects the idea that a site like this only ever compounds.

Your version: judge a competitor's profile by the top referring pages, not the total. Forums and wikis mean the links were earned and are hard to take from them. Guest posts and directories mean they were bought and can be outspent.

Google visibility is not AI recall

This is the finding the source article could not contain, because it predates the test. Start with the Google half, checked with the free Rank Tracker.

UNmiss Rank Tracker showing rtings.com at position 1 for the keyword best tv for bright room with the ranking page www.rtings.com/tv/reviews/best/bright-room, followed by reddit.com, walts.com, gagadget.com, forbes.com, star-power.com, youtube.com, avforums.com, cnet.com and avsforum.com in Google's top ten.
What to notice: position 1, above Reddit, Forbes and CNET. Read from Google at 14:58 UTC on 13 August 2026.

One honest qualification, because it changes what you should copy. The URL at #1 is /tv/reviews/best/bright-room — an editorial page from the 6,628-URL sitemap, not one of the 22,487 generated ones. We cannot show that a generated page holds a spot like this. What the generated layer demonstrably does is surround that page: it supplies the internal links into it, and it collects the citations that lift the domain carrying it.

Now the same brand through our free AI Visibility Checker, which asks AI assistants the questions a buyer asks and reports whether the brand gets named.

UNmiss AI Visibility Checker report for RTINGS showing 3 of 12 AI answers named the brand, 3 links to the site, and 1 of 2 AI platforms mentioning it. ChatGPT answering from memory scored 0 of 6 while ChatGPT Search scored 3 of 6 and linked the site as a source. Topic bars show TVs 1 of 4, Headphones 1 of 4, Monitors 1 of 2 and Speakers 0 of 2, and the cited sites list is led by rtings.com with 3, then techradar.com, whathifi.com and reddit.com.
What to notice is the split between the two rows: ChatGPT answering from memory named RTINGS in 0 of 6 answers. The same model with web search on named them in 3 of 6 and cited three of their review pages.

Everything RTINGS has built is a retrieval asset. It works brilliantly when something goes and looks, and does almost nothing when the answer comes from a model's memory. Their topic bars show the ceiling: TVs 1 of 4, headphones 1 of 4, monitors 1 of 2, speakers 0 of 2 — a category they publish in extensively, and the AI never reached for them.

One honesty note, because it cuts against the number we just published. We ran this check twice on the same day. The first run returned 1 of 12 with a citation list that did not include rtings.com at all; the run in the screenshot returned 3 of 12 with rtings.com at the top. Language models are not deterministic and a searching model gets whatever the live web hands it that minute. Treat any single run as a sample, and treat "0 of 6 from memory" as the finding that repeated.

Your version: measure both numbers on your own brand and look at the gap rather than either one. Where your Google position and your AI mention rate diverge is where the next two years of work is.

What to copy

  • The data-model-first order. Find the measurement you own before you design a template.
  • Interlinking as a property of the schema, not a task in a spreadsheet. Every generated page should link to the records it renders; every record should link back.
  • Long descriptions, normal titles. 259 characters of real summary; 60 characters of title.
  • Thousands of 810-a-month decision terms instead of one head term you will lose.
  • Machine-written facts, human-written judgement, with the line drawn deliberately.
  • Auditing by group score rather than headline score, so architecture failures surface early.

What not to copy

You probably do not have a lab. That is the load-bearing wall. Without first-party measurements the generated layer has nothing true to say, and generated pages about data you did not collect are the exact thing Google's helpful content work is aimed at.

You do not have twenty years of citations. Wikipedia links to RTINGS because RTINGS was the only place with a contrast ratio for a 2015 television. That is a decade of compounding, not a campaign — and our own chart shows it currently going backwards.

The schema becomes the ceiling. A database-driven site can only publish what its fields hold. RTINGS measures speakers and publishes speaker reviews, and the AI check still returned 0 of 2 there. An editorial site can chase a new angle in an afternoon; a programmatic one has to add a column, re-test a product line and regenerate.

And it does not buy AI visibility. If the model has to search to find you, you depend on being findable at retrieval time, which is a different discipline from ranking.

What we could not measure, stated plainly. We have no traffic, revenue or conversion data for RTINGS. Everything above is rankings, links, page counts and AI mentions — visibility, not business outcomes. Position 1 on a 810-a-month term is not proof that the model pays, and nobody outside RTINGS can verify that it does. Read this as a case study in how the visibility was built, and do not let it stand in for a profit-and-loss statement.

Run the teardown

Everything in this article was produced with free tiers, a browser and about two hours. Here is the order.

  1. Read their sitemaps first. Fetch /robots.txt, follow every Sitemap: line, and count the URLs in each. A second sitemap you did not expect is usually where the strategy is hiding.
  2. Sample twenty pages and measure the meta tags. Titles and descriptions, character counts, against the 60 and 160 display limits. Patterns show up fast, and they tell you whether tags are hand-written or generated.
  3. Audit one page. Not for the score — for which group loses points. Technical failures at scale mean the architecture is not holding.
  4. Pull the link profile. Look at the top referring pages, not the total. Forums and wikis mean earned; guest posts and directories mean bought.
  5. Check a category term. One "best X for Y" query tells you whether the model is working today, and who else is on the page.
  6. Ask the AI. Run the brand through an AI visibility check and compare it against the Google position. Where those two diverge is the opportunity.

The tools used

Every tool that produced a screenshot above, and what you actually get without paying. These limits are the ones enforced on our server, not the ones on a marketing page. We have deliberately left prices out — this category reprices constantly and a figure verified today is a liability in three months.

The jobToolFreeWhat paid adds
Check the structure holds at scaleWebsite Audit3 checks a day, no account
Find the low-difficulty buyer termsKeyword Research3 searches a day, 10 keywords per list100 keywords per list
Group them into templates and hubsKeyword Cluster Tool2 runs a day
Confirm you actually rankRank Tracker3 checks a day, 3 keywords10 keywords, competitor authority scores
See who links to you and whyBacklink Analyzer3 checks a day, 50 link rows100 link rows
Find out if AI names youAI Visibility Checker3 checks a day, 6 questions
Free, no account
Start where we started: audit one page

Step 3 of the teardown, on your own domain. In about ninety seconds you get a health score, the four group scores — on-page, technical, speed, mobile — and the named checks you failed, which is how you find out whether your architecture holds at scale before you build 22,487 pages on top of it.

  • 118 checks, the same run that scored rtings.com 90/100
  • Group scores, so you can tell copy problems from architecture problems
  • 3 checks a day, no signup, no card
Run the free Website Audit

Frequently asked questions

What is programmatic SEO?

Generating large numbers of pages from a structured data set rather than writing each one. RTINGS is the clearest example we found: 22,487 of their indexed URLs are automatically built "X vs Y" comparison pages, each rendering two complete sets of lab measurements. It works when the underlying data is real and yours; it produces thin pages when it is not.

How many pages does RTINGS.com have?

Counting both sitemaps on 13 August 2026: 6,628 URLs in the main sitemap (5,934 reviews, 265 explainers, 241 test-methodology pages) and 22,487 in a separate comparison sitemap, for roughly 29,100 in total. The biggest editorial categories are TVs at 1,235 and headphones at 1,066.

Do the auto-generated pages rank, or just the editorial ones?

We can only show the second. The page holding position 1 for "best tv for bright room" on 13 August 2026 was /tv/reviews/best/bright-room, an editorial URL. What we can show about the generated layer is that it exists at 22,487 pages, that it supplies internal links into the editorial set, and that links point at 36,960 separate pages on the domain rather than just the homepage.

Does RTINGS use AI to write its content?

Anatolii Ulitovskyi's analysis of their top eleven pages put the mix at about 25% human-written and 75% AI-generated, with the machine handling technical specification writing against structured data. We did not re-run that classification. What is checkable is that the generated layer describes measurements taken in their own lab, which is what makes it defensible.

Should meta descriptions be longer than 160 characters?

RTINGS clearly thinks so. Across our 20-page sample their descriptions averaged 259 characters, 19 of 20 ran past 160, and the longest was 697. The logic is that Google rewrites most snippets anyway, so a longer description is more raw material rather than wasted text. Note the asymmetry though: their titles averaged 53 characters, comfortably inside the normal limit.

How strong is the RTINGS backlink profile?

On 13 August 2026 our Backlink Analyzer returned 722,598 backlinks from 35,625 referring domains, Domain Trust 86/100, with 530,785 dofollow against 191,813 nofollow, and links pointing at 36,960 separate pages on the domain. The profile is currently contracting — the growth panel fell from 906,485 to 695,846 across three months.

Why does a site that ranks #1 on Google get ignored by AI?

Because they measure different things. In our check, ChatGPT answering from its own training memory named RTINGS in 0 of 6 answers, while the same model with web search on named them in 3 of 6 and cited three of their review pages. Rankings and backlinks make you findable when something searches; they do not put you in a model's unaided recall.

Can I copy this strategy without a testing lab?

Partly, and you should be clear-eyed about which part. The structure, the interlinking, the long descriptions, the bottom-funnel keyword targeting and the templated build all transfer. The reason those pages are trusted does not. If you can measure something first-hand in your own category — real prices, real delivery times, real failure rates — that is your substitute. If you cannot, you are copying the reproducible 80% and none of the defensible 20%.

How was this case study measured?

Sitemap counts and meta-tag lengths were measured directly against rtings.com on 13 August 2026; the meta-tag figures come from a 20-page sample of 14 reviews and 6 comparison pages. The audit, keyword, cluster, rank, backlink and AI visibility figures are single free-tier runs of our own tools on the same day, and each is shown in a screenshot in this article. Figures attributed to Anatolii Ulitovskyi are from his December 2025 analysis and were not re-verified. We have no traffic or revenue data for RTINGS.

The takeaway

The achievement is not a large page count. It is a measurement system that gives each comparison a useful answer.

That is the lesson to borrow: let real evidence create the pages, let people make the recommendation, and check Google visibility and AI visibility separately.

Start small: audit one page free, then decide whether your architecture can support the experience you want to build.

Measured on 13 August 2026 against rtings.com using free tiers of our own tools, plus direct sitemap and meta-tag sampling. Tool results are single runs and language models are not deterministic — the AI visibility check returned 1 of 12 on an earlier run the same day and 3 of 12 on the one shown. Figures credited to Anatolii Ulitovskyi come from his December 2025 LinkedIn analysis and were not independently re-verified. No traffic or revenue data for RTINGS was available to us. Third-party metrics change; re-run anything here before you rely on it.

Blog