Throughout 60,000+ web sites we observe every day analytics for, ChatGPT sends 8-9x the referral site visitors of the subsequent main AI-first platform, Perplexity.
Simply final month, OpenAI CEO Sam Altman revealed that greater than 800 million folks use the platform every week.
However what does it take to get talked about and cited in its responses?
Our newest analysis into top-of-the-funnel queries exhibits that not too long ago up to date “finest X” lists have been essentially the most distinguished web page sort in ChatGPT sources, together with these the place really useful manufacturers ranked themselves in first place.
There was additionally a correlation between a model positioning extremely in third-party lists and being extra more likely to characteristic in responses.
Does that conclusively imply you need to be publishing these lists your self? Earlier than I share the strategy we’re taking, let’s take a look at the info in full.
I analyzed ChatGPT responses throughout 750 top-of-the-funnel prompts in three classes: software program, merchandise, and company suggestions (e.g., ‘the very best net design companies in London’).
Whereas that quantity won’t sound like a lot, it took dozens of hours to manually categorize responses to make sure accuracy.
What I used to be searching for: How usually “finest” lists present up as citations
For software program, I used to be taking a look at phrases like “finest crm software program for enterprise” and “finest challenge administration software program”.
For merchandise, I analysed prompts round queries corresponding to “finest entry-level DSLR” and “finest dwelling treadmill”.
Within the company class, I checked out phrases associated to companies providing providers protecting search engine optimisation, social media, promoting, branding, normal advertising and marketing, and net design.
In conventional search engine optimisation, these phrases are estimated to be looked for a whole lot of 1000’s of instances per month.
That stated, it’s possible that the phrases folks enter into Google won’t be the right illustration of what’s entered in a chat platform. Sadly, AI platforms don’t present that information straight.
It’s additionally possible {that a} notable proportion of suggestions are given as the results of a multi-query dialog, moderately than a single immediate.
Because of this, I haven’t regarded into essentially the most really useful manufacturers or the top-cited domains general, however moderately into the kinds of content material sourced from these websites.
One other intention with this research is to set a benchmark for what we’re seeing throughout the board, which we will examine towards and enhance upon over time.
Sidenote.
For extra particular particulars about this analysis and its methodology, I’ve created a supplemental Google Doc that shares classes discovered and concepts for future analysis primarily based on my findings.
The desk beneath covers the most typical web page varieties we noticed throughout all three classes.
It consists of each first-party mentions—when a really useful model or product’s personal web site is cited—and third-party mentions.
The vast majority of web page varieties we analysed must be self-explanatory, however for these which might be essentially the most distinguished, right here’s some extra clarification:
- Weblog record: An inventory of suggestions in a weblog publish format, corresponding to this publish on “The 8 Greatest AI Detectors, Examined and In contrast” by our personal Ryan Legislation
- Non-Weblog Checklist: Platforms like G2 (software program) or Clutch (company), which rank suggestions however not in a weblog publish model format
- Touchdown web page: Normal pages like a selected service an company presents, or a market a SaaS firm has an answer e.g., this web page on Siege Media exhibiting they provide net design providers
- Social media: This covers websites like Reddit, Fb, Instagram, and LinkedIn
Hyperlinks could seem inline with suggestions, or as supply hyperlinks, which have been used for analysis however require a further click on to view.
Every first-party point out was manually categorized.
For third-party mentions, totaling 10,000+ particular person URLs, we used a semi-automated strategy.
We arrange dozens of handbook filters to enhance accuracy. For instance, the be taught.g2.com subdomain hosts content material varieties very totally different from normal G2 class pages.
For the remaining URLs, we categorized them with OpenAI’s GPT-5 mannequin with customized directions.
Lastly, we hung out going over every group of URLs and improved tagging the place essential, although we have been proud of 95% of what we initially noticed.
In some classes there’s a transparent shift during which web page sort featured prominently. Touchdown pages are generally cited when searching for the very best software program or company, however not often present up when researching merchandise.
Normal weblog posts on product web sites carried out much better than within the software program and company suggestions area.
For every of the classes in query, we took 250 “finest” weblog lists from ChatGPT’s responses (750 in whole) to find out the place a selected model was featured when that record was sourced alongside it.
This was all accomplished by hand, as automation wouldn’t be as correct. For instance, many lists declare they’ve a sure variety of objects, after they usually have extra and easily haven’t up to date their title or headline.
Automation to find out the place a model featured in an inventory will be much more advanced. Particularly as they’re not all the time introduced in a clear record of numbered headings.
Within the instance beneath, you possibly can see that Asana options as the highest spot in their very own record of the very best challenge administration software program:
In an analogous article on Zapier’s web site, Asana is proven because the second advice:
Each pages have been used as a supply in a ChatGPT response.
Taking all 750 lists under consideration, right here’s the distribution of the place a talked about product, model, or service supplier was featured in an related “finest” record.
Whereas there’s a transparent suggestion that being positioned extremely will increase suggestions, there’s additionally some bias right here.
Not each record has 10 or extra objects, so there’ll all the time be extra suggestions at larger record positions.
To assist remedy this, we will take a look at record distribution and what number of suggestions fall into the highest, center, or backside third of an inventory, no matter its dimension.
As we will see from the general distribution, there’s nonetheless a powerful development in direction of firms and merchandise that ranked larger within the related record.
I needed to do one higher than that and in addition share a chart centered on a separate set of lists with 10 or extra objects. There we see a much less distinguished outcome, however it’s trending in the identical course:
Another elements is likely to be in play right here, as coated within the aforementioned supplemental doc.
One fast conclusion we’d logically bounce to is that when going by content material on an internet site, LLMs aren’t parsing your entire web page, and thus solely “seeing” a number of the high suggestions.
I spoke briefly with Dan Petrovic from DEJAN search engine optimisation, who has accomplished in depth analysis on this area, and advised that may not be the case. A few of his evaluation reveals that roughly content material will be extracted relying on the area or web page being sourced.
We’re nonetheless learning this, so I hesitate to provide a definitive conclusion, however even when AI weren’t a factor, you’ll logically wish to characteristic larger on an inventory of really useful software program, merchandise or service suppliers in your trade.
I took an equal pattern of weblog lists from our three classes and checked to see after they have been revealed or final up to date. After filtering out these with out clear publish or modified dates, I used to be left with a clear record of 1,100 URLs to investigate.
79.1% have been final up to date in 2025, and 26% have been final up to date previously two months alone.
This accounts for those who have been both revealed for the primary time or later modified within the respective timeframe.
Curiously, extra articles had been up to date since publication (57.1%) than these revealed and left as is.
Preserving articles updated could assist them carry out higher in conventional search, however our information exhibits AI assistants favor even brisker content material.
Spend a while reviewing the sources ChatGPT makes use of and it’s clear they don’t seem to be all the time essentially the most legit web sites.
Many lack any human presence and look virtually solely like they have been constructed for hyperlink constructing and / or quotation functions.
Whereas the bar for high quality must be larger in the event that they’re utilizing Google and Bing as their reference sources, our personal Louise Linehan discovered that 28% of ChatGPT’s most-cited pages have zero natural visibility.
My very own analysis discovered that many questionable websites carry out effectively in Bing however not as effectively in Google.
Whereas Ahrefs Area Score isn’t an ideal indicator of web site high quality, if I take 1,000 cited finest lists throughout every class (3,000 in whole), I discovered that a good portion have been revealed on domains with low authority.
In easy phrases, they’ve fewer high-quality websites linking to their personal.
As Glenn Gabe coated in his glorious article, Google’s personal AI solutions ought to have fewer points selling legit websites, however different AI search platforms will possible have to develop stronger belief and high quality scoring going ahead.
I count on we’ll see the left aspect of this chart decline over time.
Whereas nearly all of this research was centered on ChatGPT, the prominence of those comparability articles isn’t unique to OpenAI’s platform
Actually, they have been barely extra distinguished in Google’s AI Overviews:
It’s not simply top-of-the-funnel queries the place they present up both.
Wanting on the high 1,000 most-cited pages throughout every platform through Ahrefs Model Radar, all included comparability listicles. And that’s throughout 150M+ whole prompts.
The way in which ChatGPT recommends companies and merchandise differs from the way it at the moment recommends software program.
Once you’re searching for an company, you possibly can usually click on straight on that company’s title to come out a sidebar and be taught extra about it.
Sadly, for the needs of analyzing totally different classes, this isn’t all the time the case.
As you possibly can see beneath, there was no direct web site hyperlink within the sidebar for Charle, however there was a hyperlink to their web site within the textual content beneath their title.
Sidenote.
In ChatGPT 5.1 these sidebars appear to not seem for companies, so it’s lucky that we didn’t concentrate on mentions in them.
Attributable to this inconsistency, for agency-focused prompts we determined to concentrate on the hyperlink related to a model, moderately than the hyperlink which will seem of their sidebar.
This should not have any impression on stats for issues like Weblog Lists or Non-Weblog Lists, however it could have technically allowed us to search out much more first-party homepages in our evaluation.
Whereas weblog lists have been extremely distinguished throughout all sources, when a model’s personal web site was listed as the primary advice, an inner touchdown web page was extra more likely to seem than its homepage or finest weblog record.
The desk beneath exhibits the proportion of prompts that first-party mentions confirmed up in, by web page sort.
| Software program | Company | Merchandise | |
|---|---|---|---|
| Normal Touchdown Pages | 37.2% | 30.4% | 0.4% |
| Weblog Checklist | 34% | 17.2% | 4% |
| Homepage | 15.6% | 14% | |
| Documentation (incl .PDF) | 7.6% | 0.4% | 0.8% |
| Weblog Publish | 4.8% | 0.4% | 12.4% |
| Product Web page | 0.4% | 87.2% |
Sidenote.
Observe: We aren’t anticipating columns so as to add as much as 100%, since a number of web page varieties can seem in a single response.
An easy clarification for touchdown pages outperforming weblog lists within the software program and company classes is that firms are much more more likely to have created touchdown pages protecting the providers and options they provide, than to have made a self-serving weblog record.
Although they weren’t essentially the most distinguished first-party web page sort, a self-promotional weblog record nonetheless confirmed up in additional than a 3rd of ChatGPT responses within the software program class.
I discovered they have been much more distinguished in Google SERPs,
Throughout 250 “finest X software program”-style SERPs, 169 (67.6%) featured an inventory during which the corporate writing the article ranks itself primary.
Carry out any sort of “finest [seo / web design / branding] company” search in Google and there’s a powerful likelihood you’ll see a self-promotional record rating effectively.
Within the merchandise class, it was uncommon for a corporation’s personal weblog record to seem, just because a lot of product firms don’t publish them.
You gained’t see a web page on Nike.com recommending Reebok, however they are going to publish an article on “The Greatest Winter Sneakers” and promote Nike’s personal sneakers all through.
There isn’t something new about creating pages to cowl the markets you serve and the options you provide, however you may wish to make certain these pages are priceless and as much as date.
All the information factors to a transparent reply: Businesses and SaaS firms ought to publish self-promotional “finest” lists.
Numerous well-liked manufacturers like Shopify, Slack, Salesforce, HubSpot, and Superside have written them, and there seems to be no draw back from a rating or quotation perspective.
There’s additionally an argument for being pleased with what you create, and for serving to guests navigate extremely saturated markets.
That stated, information is just one a part of the equation.
Throughout my very own analysis, I’ve usually been pissed off by the lists I come throughout. Many are overly self-promotional, or just don’t hyperlink out to the options they advocate.
Whereas his feedback have been aimed toward creating unhealthy person experiences basically, moderately than focusing on self-serving weblog lists, Mark Williams-Cook dinner made an awesome level on LinkedIn that poor person indicators may meet up with you in the end.
Siege Media founder Ross Hudgens agrees these indicators are necessary, however argues you possibly can place your individual model first and nonetheless present worth.
Company proprietor Wil Reynolds, whom I’ve revered for years, had one other opinion:
Any marketer price a rattling is aware of itemizing your self on the high of an inventory hosted in your web site doesn’t construct belief with consumers.
It wasn’t a standard incidence, however I did discover examples of firms that includes decrease down their very own lists.
DoorLoop positioned itself seventh on this record of rental administration software program suppliers
Whereas it’s admirable as somebody with no connection to the enterprise, there’s additionally the argument that they’re simply serving to opponents with their very own advertising and marketing.
You’ll should resolve which strategy is smart for what you are promoting, however let me at the least share our present take.
At Ahrefs, our considering is straightforward: We could publish extra comparability posts as we roll out new options or enter new markets, however they’ll proceed to make up lower than half a % of our whole weblog content material.
We’re not going to create them for each software or potential use case.
We gained’t all the time put ourselves within the high spot, as is already the case.
We’ll proceed to make it clear that we are the Ahrefs being really useful.
(Nearly each firm selling itself does so within the third particular person, as in the event that they’re subtly making an attempt to miss that they’re writing about their very own enterprise.)
And we’ll maintain linking on to various options so guests can analysis them on their personal.
One change we are making because of this analysis is to rejoice extra of our wins.
There have been many instances the place manufacturers have been talked about alongside constructive firm updates, as proven beneath for Monday.com.
In one other case, when fitness center administration software program Kilo was talked about, it was alongside their announcement of incomes an award from Capterra.
We may do a significantly better job there, with the potential added advantage of giving LLMs attractive commentary to make use of with any mentions.
Our main focus will proceed to be on issues which have naturally gotten us talked about on the lists of others, like always updating our product with new options, sharing authentic information and analysis, and being lively the place our viewers is.
Mentions from manufacturers like Zapier are all the time good to see
As we’ve accomplished after we spent $400,000 operating an occasion, $1M+ sponsoring creators, or revealing how we resolve the enterprise potential of latest weblog posts, we’ll maintain you within the loop about how our strategy to AI search visibility evolves.
Keep tuned for some behind-the-scenes updates I’m engaged on there.
When you’ve got any ideas on how I may enhance the subsequent iteration of this research or the rest you prefer to me to dive into, please let me know over on LinkedIn or X.
