More searches now end on a summary instead of a click, and the summary repeats whatever its sources got wrong in the same confident voice. Use the answer to find the source, not as the source.
- A summary inherits its sources' mistakes. It cannot tell a well-sourced page from a well-optimised one, and it sounds just as sure either way.
- Click through when it matters. Health, law, money, safety, specific people, anything where the date matters or the question is contested, and anything you plan to repeat.
- Search for the original. Use the
site:operator, phrase the query the way a practitioner would, and add the year when recency matters. - Open the citation. A cited answer is not verified until somebody opens the link.
- Brands now work to be named in answers. Treat a product an answer engine names as a lead to check, not a neutral recommendation.
The prices here come from a third-party survey, not from the vendors’ own pages, and were correct when this site read it on 11 Sep 2026. Prices change often and this site no longer updates them, so check each vendor’s own site before you rely on one. Survey read: honeyb.ai, AI visibility software pricing.
An answer engine reads pages and summarises them. So it inherits whatever those pages got wrong, it cannot distinguish a well-sourced page from a well-optimised one, and it presents the result in the same confident register either way. Nothing about the fluency of an answer tells you anything about its accuracy — and the step where you noticed the source is now optional.Measured context, both read at source Sep 2026. Zero-click: SparkToro on Similarweb clickstream data, via Search Engine Land — "Google searches ended without a click 68.01% of the time in the U.S. during the first four months of 2026 … That's up from 60.45% in 2024, a 7.56-point increase in two years." Machine-written articles: Graphite — "The quantity of AI-generated articles has surpassed the quantity of human-written articles being published on the web," and "The proportion of AI-generated articles has plateaued since May 2024." See also AI and the open web.
When the answer is enough — and when it is not
Usually fine
- Settled facts with no consequence: conversions, definitions, dates long past, how a common thing works.
- Getting oriented in a topic you know nothing about, before you go and read properly.
- Finding the name of the thing you are looking for, so you can then search for it directly.
- Anything where being wrong costs you nothing.
Click through, every time
- Anything that changes what you do — medical, legal, financial, safety.
- Anything about a specific person or organisation. Summaries blur individuals together with astonishing confidence.
- Anything where the date matters. An answer engine rarely tells you how old its underlying source is.
- Anything contested. A summary of a live disagreement usually reads as though the argument is over.
- Anything you intend to repeat as fact.
Let a summary handle small, settled questions. Click through whenever the answer would change what you do, is about a specific person, or depends on a date.
Habits that still surface real things
Search for the source, not the description of it
Most pages about a report are worse than the report. Go for the thing itself: the agency, the paper, the official filing.
internet crime report 2025 site:ic3.govThe site: operator limits results to one website, which is often the quickest way to reach the original document.
Search the way a practitioner would phrase it
General questions are exactly what mass-produced pages are optimised to answer. Specific, awkward phrasing — the words someone would use who has actually done the thing — reaches different pages entirely.
why does my sourdough starter smell like acetone day 4Rather than "sourdough starter problems". The odd detail is what filters out the generic.
Put the year in when recency matters
Answer engines are frequently summarising material older than they imply, and older pages accumulate authority. If the answer would be different last year, say the year.
EU AI Act transparency obligations 2026Follow the citation, do not just note that one exists
A cited answer feels verified. It is not verified until somebody opens the link. Citations are sometimes attached to the wrong claim, sometimes to a page that merely repeats it. On this page's reasoning, opening one link is the habit that matters most.
Ask the answer engine to argue with itself
"What is the strongest case against that?" or "What would someone who disagrees say?" turns a confident summary into a visible disagreement — which is usually a truer picture of the evidence, and it costs one extra sentence.
Keep a short list of places you go directly
The most valuable defence against ranked results is not needing them. Three or four sources you trust in the areas you care about — bookmarked, or followed by RSS — will beat search for anything you care about repeatedly.This site publishes a feed for exactly that reason: follow the record (published 22 Aug 2026)
Go to the original: search for the document itself, add the year when it matters, and open at least one cited link before you trust an answer.
What is now being optimised at you
Search optimisation aimed at ranking in a list. Its successor aims at being quoted inside the answer — an industry has formed around getting brands cited by AI systems, sold under names like answer engine optimisation and generative engine optimisation, with monitoring tools sold at every price point — from free checkers and about $29 a month at entry level up to enterprise contracts in the thousands.Pricing survey read at source 11 Sep 2026 — honeyb.ai, "AI Visibility Software Pricing: What the Tools Cost in 2026": "The prices for that read span two orders of magnitude, from a free checker to enterprise contracts in the thousands per month." And: "Self-serve tools publish a number you can act on in an afternoon, and they cluster between $29 and roughly $295 a month."
Two things follow, and they point in opposite directions.
The unsettling one: when an answer engine names a product, a tool or a company, somebody may have worked deliberately to be the name it reaches for. That influence is invisible in a way a sponsored link never was — there is no label, and the recommendation arrives in the same neutral voice as everything else.
The reassuring one: the tactics that actually work are largely the tactics of being useful — clear structure, direct answers, accurate facts, visible sources, corroboration elsewhere. Machines reward legibility, so this optimisation is far less corrosive than the keyword stuffing that preceded it.Terminology and industry state recorded Aug 2026, pricing rechecked Sep 2026. Corrected Sep 2026: this page said monitoring tools are "priced in the hundreds of dollars a month". A pricing survey read at source (honeyb.ai) puts it wider — "The prices for that read span two orders of magnitude, from a free checker to enterprise contracts in the thousands per month." The same practice, examined from the publishing side, on Pages Written For Machines, Read By No One — including this site's own use of it.
The habit that covers it is the one already on this page: follow the citation. Optimisation can influence which source gets named. It cannot make that source say something it does not say.
When an answer engine names a product, tool or company, treat it as a lead rather than a recommendation, and open the source to see what it actually says.
Use the answer to find the source; do not use the answer as the source. Everything above is a way of doing that faster.
Answer engines are genuinely useful and this site is built to be readable by them. The problem is not summarisation — it is that summarisation removed the moment where you saw who was talking, and that moment was doing more work than anyone realised. It is where you noticed a source was a vendor, or an outlet you distrust, or a page written last decade. Getting that moment back costs one click, and it is the whole difference between being informed and being answered.