Most trend articles are a list of things the author saw on Dribbble. This one is built from
our own numbers: what happened to a real agency site over the past year, what we measured, and
what we changed because of it. Where we do not have data, we say so.
Three things genuinely changed in 2026. Assistants started sending real traffic, the cost of
publishing collapsed while the value of publishing volume collapsed with it, and the parts of a
site nobody looks at started deciding whether the rest of it gets found at all.
1. AI assistants are now a referral channel, and your analytics is probably miscounting it
On our own site, ChatGPT sent 46 visits and one enquiry over 180 days. That is small. What
matters is the second number: Google Analytics’ built-in “AI Assistant” channel counted 32 of
those 46. The other 14 were filed as Referral or Unassigned, and the single enquiry landed in the
misfiled group.
So the report you would naturally trust showed AI traffic as converting zero times, when it
had actually produced a lead. If you are judging whether assistants matter by looking at that
channel, you are reading an undercount.
The fix is not a tool. Match the hostnames yourself: chatgpt.com,
perplexity.ai, gemini.google.com, copilot.microsoft.com,
claude.ai. Match on the exact host or a subdomain of it, never a substring. Our own
property contains a spam referrer literally called chatgpt.comvvvvvvvv…, and a
substring match counts it as real traffic.
The part people miss: your server log already knows
Referrals only count people who clicked. Far more often, an assistant reads your page and
answers without sending anyone. That still leaves a trace, in your access log, and the useful
distinction is what each crawler is for:
- User-triggered (
ChatGPT-User,Perplexity-User,
Claude-User): a person asked a question and the assistant fetched your page to
answer it. This is the closest thing to evidence that you were part of an answer, and it costs
nothing to measure. - Search index (
OAI-SearchBot,PerplexityBot,
Claude-SearchBot): keeping a current copy so you are eligible to be cited. - Training (
GPTBot,ClaudeBot,
CCBot,Google-Extended): a copy taken to train a model. This tells you
nothing about whether anyone sees you, and whether to allow it is a policy decision rather than
an SEO one.
Before you go looking, check that you are recording it. Bitnami’s Apache ships with
CustomLog "logs/access_log" common, and the common format contains no
user agent field at all. Ours did not, which meant no AI crawler had ever been logged and none of
that history could be recovered. If your log lines end at the response size, you are in the same
position:
# Bitnami default: no user agent, so no AI crawler is ever recorded
LogFormat "%h %l %u %t \"%r\" %>s %b" common
# What you want instead. The X-Forwarded-For matters behind a CDN,
# where %h is the edge IP rather than the visitor.
LogFormat "%h %l %u %t \"%r\" %>s %b \"%{Referer}i\" \"%{User-Agent}i\" \"%{X-Forwarded-For}i\"" withagent
CustomLog "logs/access_log" withagentAdd it as a new format nickname rather than editing the stock common line, so a
platform update that rewrites the defaults cannot quietly switch you back.
2. Being cited is a different job from ranking, and mostly a writing problem
We asked a model the questions our actual buyers would ask. “Which agencies offer white label
WordPress development for other agencies?” “Recommend a white label web development partner in
Europe.” We were named in none of them. It answered with 10up, Human Made, WebDevStudios,
Netguru, Ciklum, Acquia and Lullabot instead.
That is a more useful result than a score. It tells you who the assistant currently considers
the answer, which is the competitive set you are actually up against.
One caveat worth stating, because the tools selling this rarely do: a model without live web
search is telling you what is in its training data, not what it cites today. Every “AI visibility
score” on the market is produced the same way, by asking repeatedly and counting. There is no API
from OpenAI, Anthropic, Google or Perplexity that reports how often they showed your site.
What appears to work is unglamorous. Assistants quote passages that answer one question
precisely and show their working. That means the exact error message, the version number it
applies to, the configuration that fixes it, and the case where the fix does not apply. A page
that hedges is a page with nothing quotable in it.
3. Publishing volume stopped working, and we have our own receipts
The going rate for automated content is around 30 articles a month. We can tell you what
volume did on our own site, because we ran that experiment before it was fashionable: 33 service
pages produced three non-brand clicks in six months. Sixteen blog posts produced
two.
Meanwhile the single best topical signal on the whole site was one post answering an exact
WooCommerce question, drawing twelve related searches on its own. It had zero internal links
pointing at it. Nothing on the site linked to the one page Google half-believed in.
The lesson is not “write less”. It is that a page earns nothing until something links to it and
it answers a question better than the three results above it. Volume makes both of those harder,
because every new page splits attention and internal links further.
4. The internal link graph is where sites quietly fail
This is the finding that surprised us most, and it is invisible in Search Console by design.
Search Console reports what Google did, never why it declined.
We published three technical reference pages. They returned HTTP 200. They carried correct
canonical tags. They were in the sitemap, and the sitemap was submitted. Google reported
URL is unknown to Google for all three, for days.
Nothing linked to them. A page reachable only from your sitemap has the weakest possible claim
on being crawled, and no amount of resubmitting changes that. When we crawled our own site and
counted inbound links properly, six pages had none at all and thirty-nine hung off a single link,
usually a menu entry.
If you do one thing from this article, crawl your own sitemap and count inbound links per URL.
It takes an afternoon and it finds problems no rank tracker reports.
5. A trend we are not endorsing: automated link networks
The same platforms selling 30 articles a month now bundle “contextual backlinks” from networks
of a thousand or more sites. For an agency doing white-label work this is worse than it looks: if
a client site is penalised for participating, the penalty is on their domain and your name is on
the invoice. You cannot unwind someone else’s link profile.
We mention it because it is genuinely the fastest-growing offer in this space, and declining it
is a position, not an oversight.
What we would actually do in 2026
- Check your log format before anything else. If it does not record the user
agent, you have no AI data and cannot get the past back. - Stop trusting the AI channel in Analytics. Match assistant hostnames
yourself, exact host or subdomain, and look at conversions per platform. - Crawl your own site and count inbound internal links. Fix the pages that
already earn impressions and have none. That is the cheapest work available to most sites. - Write fewer, more specific pages. One exact answer with the error message,
the version and the failure case beats ten overviews, for Google and for assistants. - Track which page assistants send people to. Ours send them overwhelmingly to
one page. That is the page worth strengthening, and we would not have guessed which one.
None of this is a prediction. It is a list of things we measured on our own site this year,
including the parts that were embarrassing. If you want to compare notes on what your own log and
link graph say, we are happy to look.
