I build great links for agencies & in-house marketers at @loganix
Vancouver, BC
Joined September 2011
- Tweets14.3K
- Following466
- Followers3.6K
- Likes10.8K
In the land of link building:
- If it's too good to be true, it is.
- You get what you pay for, mostly.
- Pricing by DR is how you end up paying premium for a site that ranks for nothing
...only yesterday I reviewed a truly awful link profile built by a blog exchange company. The customer had no idea and was very happy with their work.
Adam at Loganix retweeted
@elder_plinius just leaked opus 5's system prompt
it's 198kb, 2,049 lines, and over 200,000 characters
i read the whole thing. here's what matters for seo and ai search:
1/ the prompt states that claude's "reliable knowledge cutoff, past which claude can't answer reliably, is the end of may 2026"
fable's was the end of january
so events through february, march, april, and most of may went from being found only through retrieval-only to potentially included in the training data
2/ the retrieval instructions are close to identical to fable's
same instruction to search for "questions that appear historical or settled but are phrased in the present tense"
same year-dating correction, that "latest iphone 2025" when the year is 2026 returns stale results
same carve-out telling claude not to search for "knowledge questions, shopping recommendations, general advice"
when i wrote up fable in january, my position was that i was describing one model's behavior and it might not survive the next release
this leak proves it did in fact survive the next release. word for word in places
so this is looking like it's anthropic's standing retrieval policy (at least for now)
3/ there's a literal trigger word list in there
the prompt names "current" and "still" as signals
version-style names too, because "partial recognition isn't current knowledge"
these are the seemingly insufficient words that flip a query from the training data to live retrieval
4/ the prompt scales how hard claude looks
"1 for a single fact; 3-8 for medium tasks; 8-20 for deeper or broader questions"
and when a query names multiple things, claude is told to "search for each one separately rather than combining them into one query" because a combined query returns surface-level results for all of them
so, the "best x vs y vs z" query doesn't run a single search across all entities mentioned in the prompt. instead, it fires a separate search for each entity and compiles what was found across multiple resources into a single response
5/ web search is one of five ways claude can answer, and it's the last one it reaches for
the full system prompt gives claude the training data, a persistent memory filesystem, search over the user's own past chats, mcp connectors, and then the web
the memory block is interesting
claude gets per-user files it "re-reads at the start of every conversation," and it's instructed to write to them mid-conversation without being asked
a user says once that they use a tool, that goes in a file, and the file comes back every session
no retrieval step here, which means brand recall inside the user's own history is now a distribution channel
and that about wraps it up. hope you found that helpful. peace!
teapawt.org/loganix/status/2065280…
@elder_plinius just leaked fable 5's system prompt
it's 120k characters long (fyi: pliny's leak is an excerpt)
we read it. here's what matters for seo and ai search:
1/ claude is instructed that it "defaults to searching for questions that appear historical or settled but are phrased in the present tense"
meaning that if someone asks "what's the best link building service," claude searches for it live. even if it already has a trained opinion
present tense = live retrieval
2/ the prompt states that "claude's reliable knowledge cutoff, past which claude can't answer reliably, is the end of jan 2026"
that date is the dividing line for your geo strategy
before it: claude may or may not search depending on the query
present-tense questions, year-dated queries, anything that signals freshness, those trigger live retrieval
knowledge questions, shopping recommendations, general advice, those get answered from training data
after it: claude has no choice but to search. it doesn't know what happened. live retrieval becomes the default
so content published after january 2026 is only accessible to claude via live search, which means everything we talk about with freshness signals, present-tense framing, and year-dating your content matters even more for anything recent
3/ the prompt tells claude that "'latest iphone 2025' when the year is 2026 returns stale results" and instructs it to correct for this
so claude is actively building queries with the current year in mind. a page titled "best X 2026" is matching how the model searches, not just signaling freshness
keep your content dated
4/ the prompt instructs claude to "search_mcp_registry first. a connector is one click to connect — always better than browsing. browser only after search comes back without it."
it also explicitly tells claude not to search for "knowledge questions, shopping recommendations, general advice."
two separate ways the web never gets consulted
for product and shopping queries, claude answers from training data. not live retrieval. that means you're competing on third-party coverage, brand mentions, and authority signals built before the jan 2026 training cutoff, not on fresh content
for tool businesses: build an mcp connector and you bypass retrieval entirely for users who have it installed. become the pipe
for service businesses (us included): you can't directly engineer the connector layer without building a tool. but understanding that web search isn't always in the picture changes how you think about geo strategy entirely
5/ the prompt instructs claude to direct users to "the national alliance for eating disorders helpline instead of NEDA, because NEDA has been permanently disconnected"
what this tells us is that the model has named resource preferences that can override retrieval entirely. how far that extends beyond safety-critical resources, we genuinely don't know
but the fact that it exists at all changes how you think about what it means to be the go-to answer in your category
lots to digest here. peace!
Adam at Loganix retweeted
there's a debate in seo that never really gets settled: once google catches a site running parasite seo, does it stay watched, or does the penalty just... end?
i pulled two and a half years of traffic and page-count history on a major indian news site, one google already caught and penalized once, to find out
found something that answers it. dropping the full story here before the end of the week, so keep an eye out
i just uncovered a massive parasite seo operation running across multiple freelance platforms
gambling networks, spam rings, ai content farms, all disguised as freelancer profiles
one example: an illegal indian gambling operation using freelancer(dot)com as the host, pulling over 900k monthly visits before google caught it
this is one of hundreds, potentially thousands of parasites
dropping the full story same time tomorrow, so keep an eye out
An Indian betting operation parasite that was pulling 4.2M monthly organic visitors!!!
Gambling site-builds, adult content bait, a bot that "predicts" casino outcomes, all live, all ranking, all making money from their hosts, Upwork, Fiverr, Freelancer.
Absolute must read:
Adam at Loganix retweeted
our july 4th sale is now live!
use code USA250 to get 10% off link building, ai citations, and local SEO
runs runs through to july 17
One of the more interesting things happening in AI search:
The URL path is becoming a strong trust signal.
Put content in /press-releases/ and AI systems classify it as self-promotion.
But put the same content in /news/ and now it is trusted by AI models across the board.
This is shaping retrieval, so we built our newest product around it.
we just launched our newest product, AI Press Release
the product came out of a simple test:
we sent the same press release to 8 different site subsections, then tracked which pages appeared in ai search results across chatgpt, gemini, perplexity, and google ai overviews
one subsection consistently surfaced
seven frequently didn’t
here’s the pattern we found:
the paths split into two clear outcomes
yahoo finance /news/ and business insider /markets/ generated repeated ai citations across multiple platforms
meanwhile, /press-releases/ sections on fortune, reuters, and similar outlets rarely surfaced for the same query sets
the content was the same, and the authority tier was similar
the only variable was where the content lived on the site
why would the subsection matter?
our testing suggests ai retrieval systems use publication context as a trust signal
content at /press-releases/ reads as promotional and self-reported, exactly the type of content ai search platforms have started to filter out
content in a /news/ section carries an editorial signal, aka something someone is saying about you, not what you’re saying about yourself
this is why it surfaces as a cited source
the path is read before the content
that’s the new nature of ai search platforms
they trust third parties’ word (sometimes your competitors') over yours, which makes it hard for you to control the narrative
but here’s where you can turn that on its head: an ai press release placement.
just one placement does two things
retrieval layer: ai searches the web in real time. yahoo finance /news/ surfaces in that retrieval. the release gets cited
base model layer: llms are trained on large corpora. brand mentions on authoritative domains feed future training cycles
each placement builds the model's awareness of the brand over time
every placement we tracked achieved at least one ai citation within 7 days for its target query set
that’s not a warranty, but an observed outcome from the test
what we control is the mechanism: placement selection, subsection targeting, formatting, and distribution structure
that’s the idea behind our newest service, ai press release
research announcement content that’s placed on editorial-style news subsections and distributed through the prnewswire cascade
built for ai visibility from format through placement context
use aiprlaunch for 10% off your first order
expires june 20
read more here: loganix.com/introducing-ai-p…
Adam at Loganix retweeted
google published guidance on how to optimize for ai search
a lot of the seo community took it as "geo is just seo, case closed"
here's what isn't being talked about: google has a financial incentive to keep your optimization dollars inside their ecosystem
their guide only covers their own ai search platforms, ai overviews and ai mode
it doesn't touch on chatgpt, perplexity, or claude, where 83% of ai usage happens
@myeyesshine_ maps the evidence here: teapawt.org/myeyesshine_/status/20…
Adam at Loganix retweeted
the more you spend on a link, the less you pay per DR unit
DR 20-29: $7.35 per DR unit
DR 60-69: $3.04 per DR unit
buying lower DR to save money costs more per unit of authority
we dug into 2,443 real orders from our own marketplace to find out what buyers are actually getting wrong
full study: teapawt.org/loganix/status/2052057…
Adam at Loganix retweeted
"good seo is good geo."
sure, to a point.
but when you're building a press placement, a directory listing, a brand mention, are you only thinking about what it does for google?
because, love it or hate it, it's no longer just doing something for google.
a press placement is a backlink for google. a training signal for chatgpt. and a brand entity signal for gemini, all at the same time.
if you're only thinking about what it does for your rankings on google, you're missing what it's doing for the rest.
there's no point continuing to pretend that search and information retrieval is just a google thing.
let's face it, there are now multiple major retrieval systems your audience uses to find information.
google, chatgpt, perplexity, gemini, claude, and more.
speaking generally, they all pull from the same web. they may even "read" the same sources. the difference here is that they can and do extract different things from them.
so why aren't you optimising for all of them at once?
every tactic can serve multiple systems simultaneously if you're thinking about it the right way.
in doing so, you're hedging across platforms. you're future-proofing your brand as ai search keeps growing. and you're capturing value right now, not just banking on where google traffic is today.
it's time to stop thinking of optimization as a google-only discipline.
with just a few adjustments to how you approach it, you can be building for multiple platforms at once and hedging your bets against whatever the future brings.
so good seo really isn't good geo
here's a deeper look at why: teapawt.org/myeyesshine_/status/20…
The page-targeting map is the part I think most people skip when planning link campaigns.
We see it all the time in client profiles: 40 guest posts all pointing at the homepage with brand anchors.
Or the opposite. Every link pointed at service pages with exact-match anchors.
Both are missing the point.
Brand links go to the homepage. That’s what they do.
Guest posts go to service pages. That’s what they’re for.
Mixing those up wastes budget and looks unnatural to both algorithms we've seen the source code for from leaks.
the biggest mistake link-building mistake? being sold on dr/da
the 2024 google api leak confirmed google doesn't use either metric
here's what google actually uses:
dr and da are third-party approximations
neither sees real traffic data, user behavior, or the spam signals google uses
thankfully, though, google’s black box, exposed through the api leak, has been cracked, giving us the opportunity to really assess the quality of a backlink
here are the strongest attributes google uses to assess backlinks:
1. siteauthority: google's own domain-level trust score. the higher a site's siteauthority, the more trust and ranking power a link from that site carries
2. pagerank: the base authority signal. how much ranking power flows through a link is determined by the authority of the linking page and by how many other links share that authority
3. titlepagerank: pagerank calculated specifically for the main content area of a page. a link placed editorially in the body of a relevant article carries more weight than one in a footer, sidebar, or navigation
4. homepagepagerankns: the pagerank of the linking site's homepage. even when a link comes from a deep inner page, google factors in the homepage's authority into its evaluation
5. sitefocusscore: how topically coherent the linking site is. a focused niche site scores differently from one that covers everything and nothing
6. pagerankweight: the actual per-link weight fed into google's pagerank calculation. not every link from the same page passes the same weight. a link buried amongst dozens of others on the same page is worth less than one sitting on a page with few outbound links
7. parallellinks: the number of additional links from the same source page pointing to the same target domain. the more links a page sends to your domain, the less each individual one passes
8. firstseendate / lastupdatetimestamp: google tracks when a link was first seen and when it was last updated, confirming link age and recency are tracked signals
9. topicembeddingsversioneddata: scores the topical and semantic relevance between the linking site and your page. a link from a site operating in the same topic space carries more weight than one from an unrelated domain
10. anchormismatchdemotion: a confirmed demotion signal. if the anchor text doesn't match the content of the target page, the link can actively work against you
11. scamness: a scam detection score ranging from 0 to 1023, applied directly in google's ranking system. a link from a site with a high scamness score carries those signals with it. dr has no equivalent signal
and it's not just the google leak saying this
the 2023 yandex source code — the most complete look at a major search engine's ranking infrastructure ever made public — shows the same pattern:
yandex deprecated their entire classic link analysis module. all 60 factors. gone
they were replaced by browserpagerank: link signals validated by actual user behavior
a link from a page with no traffic generates no behavioral data, which means it generates no modern link signal to pass
the penalty coefficient for link schemes in yandex's code (−0.181) almost exactly mirrors the positive weight of pagerank (+0.183)
the system is architecturally designed to cancel out what you gain from buying the wrong links at the exact same magnitude they'd help if you earned the right ones
then there's the doj antitrust case
google's vp of search testified under oath that navboost — their behavioral click-signal system — is more powerful than any of their deep learning models for ranking
the internal document entered as evidence: "learning from logs is the main mechanism behind ranking."
three independent sources. same conclusion
knowing what signals matter is one thing
knowing which link type produces which signals — and which pages each type should point to — is the fuller picture
we mapped the full spectrum this week: every link type, what each one does, where it fits, and what the leaked data shows about how they're actually weighted
link to the guide in the comments
Reddit is a hard nut to crack:
1. Posting links is outlawed in most subreddits
2. High-intent subreddits require months of karma building
3. Most subreddits hate/ban self-promotion
You can't buy your way in. You have to earn it... or work with someone who already has.
google paid reddit $60m/year for its data
within months, reddit’s search visibility jumped 342%
it’s now the #2 most visible domain in US google search
here’s what that means for seo, AI search, and b2b marketing 🧵
Adam at Loganix retweeted
currently, building coverage across the pages AI retrieves is still achievable without huge budgets
that won’t last
early movers compound
once your brand shows up consistently across those pages, displacement by gets a lot harder
strategy guide: loganix.com/backlinks-for-ai…
Adam at Loganix retweeted
AI citations are probabilistic, not positional
70% of content reshuffles per regeneration
target: appearing in 40%+ of regenerations for your category queries
low competition: 3-4 placements. moderate: 5-7. high: 8-12+
full targeting framework: loganix.com/backlinks-for-ai…
Adam at Loganix retweeted
getting a brand cited in AI search isn’t by chance
that's why we built our new AI visibility service, AI Brand Links, with:
query research built in
site selection built in
content engineering built in
competitive positioning built in
how it works:
loganix.com/backlinks-for-ai…
Adam at Loganix retweeted
chatgpt has taken a leaf out of the google playbook, and is now punishing/deprioritizing low-effort listicle spam
@lilyraynyc
